Professional Documents
Culture Documents
Dynamic Syntax
Dynamic Syntax
Preface ix
1 Towards a Syntactic Model of Interpretation 1
1.1 Natural Language as a Formal Language? . . . . . . . . . . . . . 1
1.2 Underspecication in Language Processing . . . . . . . . . . . . . 3
1.3 The Representational Theory of Mind . . . . . . . . . . . . . . . 7
1.4 Pronominal Anaphora: Semantic Problems . . . . . . . . . . . . 9
1.4.1 The Problem of Multiple Ambiguity . . . . . . . . . . . . 9
1.4.2 The Problem of Uniqueness . . . . . . . . . . . . . . . . . 12
1.4.3 The Problem of Indirect Reference . . . . . . . . . . . . . 13
1.4.4 Quantication . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.5 Syntactic Processes of Anaphora . . . . . . . . . . . . . . 15
1.5 The Anaphora Solution { Towards a Representational Account . 16
2 The General Framework 18
2.1 A Preliminary Sketch . . . . . . . . . . . . .. . .. .. .. . .. 19
2.2 The Data Structures of the Parsing Model .. . .. .. .. . .. 27
2.2.1 Atomic Formulae . . . . . . . . . . . .. . .. .. .. . .. 31
2.2.2 Tree Modalities . . . . . . . . . . . . .. . .. .. .. . .. 36
2.2.3 Basic Tree Structures . . . . . . . . .. . .. .. .. . .. 39
2.2.4 Partial Tree Structures . . . . . . . . .. . .. .. .. . .. 41
2.2.5 Requirements . . . . . . . . . . . . . .. . .. .. .. . .. 42
2.2.6 Descriptions of Tree Structures . . . .. . .. .. .. . .. 49
3 The Dynamics of Tree Building 55
3.1 The Parsing Process { A Sketch . . . . . . . . . . . . . . . . . . . 56
3.1.1 A Basic Example . . . . . . . . . . . . . . . . . . . . . . . 56
3.1.2 A Left-Dislocation Example . . . . . . . . . . . . . . . . . 64
3.1.3 Verb-nal Languages and the Grammar-parser Problem . 67
3.2 The Parsing Process Dened . . . . . . . . . . . . . . . . . . . . 76
3.2.1 Computational Rules . . . . . . . . . . . . . . . . . . . . . 80
3.2.2 Lexical Transitions . . . . . . . . . . . . . . . . . . . . . . 89
3.2.3 Pragmatic Actions and Lexical Constraints . . . . . . . . 95
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
iv
Contents v
4 Linked Tree Structures 103
4.1 Relative Clauses { Preliminaries . . . . . . . . . . .. .. . .. . 103
4.1.1 The LINK Relation . . . . . . . . . . . . . .. .. . .. . 104
4.1.2 The Data Reviewed . . . . . . . . . . . . . .. .. . .. . 105
4.2 The Analysis { A Sketch for English . . . . . . . . .. .. . .. . 109
4.2.1 Dening Linked Tree Structures . . . . . . .. .. . .. . 110
4.2.2 Relativizers Annotating Unxed Nodes . . .. .. . .. . 111
4.3 Relatives: Towards a Dynamic Typology . . . . . . .. .. . .. . 121
4.3.1 Relativizers Projecting a Requirement . . . .. .. . .. . 121
4.3.2 Variation in Locality . . . . . . . . . . . . . .. .. . .. . 128
4.3.3 Topic Structures and Relatives . . . . . . . .. .. . .. . 130
4.3.4 Variation in Order { Head-nal Relatives . .. .. . .. . 133
4.3.5 Head-internal Relatives . . . . . . . . . . . .. .. . .. . 139
4.3.6 The Potential for Lexical Variation . . . . . .. .. . .. . 142
4.4 Genitive Constructions as LINK Structures . . . . .. .. . .. . 144
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . .. .. . .. . 148
5 Wh Questions: A General Perspective 150
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . 150
5.2 The Semantic Diversity of wh Questions . . . . . . . . . . . .. . 151
5.2.1 Scopal Properties of wh Expressions . . . . . . . . . .. . 154
5.3 Wh -initial vs wh -in-situ Structures . . . . . . . . . . . . . . .. . 156
5.3.1 Wh -in-situ Structures . . . . . . . . . . . . . . . . . .. . 158
5.3.2 Wh -in-situ from a Dynamic Perspective . . . . . . . .. . 161
5.4 Expletive wh Structures . . . . . . . . . . . . . . . . . . . . .. . 163
5.4.1 Partial Movement . . . . . . . . . . . . . . . . . . . .. . 163
5.4.2 Partial Movement as a Re
ex of a Requirement . . . .. . 169
5.5 Wh Expressions and Scope Eects . . . . . . . . . . . . . . .. . 186
6 Crossover Phenomena 190
6.1 Crossover { The Problem . . . . . . . . . . . . .. .. .. . .. . 190
6.2 Crossover { The Dynamic Account . . . . . . . .. .. .. . .. . 196
6.2.1 Crossover in Relatives . . . . . . . . . . .. .. .. . .. . 196
6.2.2 Crossover Phenomena in Questions . . . .. .. .. . .. . 213
6.3 Summary . . . . . . . . . . . . . . . . . . . . . .. .. .. . .. . 221
7 Quantication Preliminaries 223
7.1 Introduction . . . . . . . . . . . . . . . . . .. . .. .. .. . .. . 223
7.2 Scope Eects and Indenites . . . . . . . .. . .. .. .. . .. . 224
7.3 Quantication . . . . . . . . . . . . . . . . .. . .. .. .. . .. . 231
7.3.1 Quantied NPs . . . . . . . . . . . .. . .. .. .. . .. . 234
7.3.2 Scope . . . . . . . . . . . . . . . . .. . .. .. .. . .. . 239
7.3.3 Term Reconstructions . . . . . . . .. . .. .. .. . .. . 245
7.3.4 Applications { E-type Anaphora . .. . .. .. .. . .. . 249
vi Contents
Ruth Kempson
Wilfried Meyer-Viol
Dov Gabbay
King's College London, July 2000
1
1
2 Towards a Syntactic Model of Interpretation
John ? John ?
Upset ?
(c) P AST :? (d) P AST : Upset(Mary )(John)
indices, but also relations dened on those entities, the situation providing the
interpretation for the sentence John had a heart attack right outside the hospital
accordingly, having not only the entities John, a heart attack and a (unique)
witness satisfying the expression the hospital, but also the relations between
hospitals, doctors and medical conditions needing treatment, on the one hand,
and between heart attacks and medical conditions needing treatment, on the
other. But if such relations are dened as part of the situation satisfying John
had a heart attack right outside the hospital, then the situation becomes quite
unlike the very narrow concept of situation required to yield the right result for
the earlier coat-choosing case.
Evidence that no extension of the notion of a situation to incorporate entities
is suÆcient for a complete characterization of anaphora is presented by cases in
which the situation described may not include an individual, and yet reference
to it by a pronoun can be successful. Consider (1.32) uttered to a friend who
has just split up from his partner and is paranoid about meeting her at every
turn:
(1.32) I'm having a party. Don't worry: she won't be there.
The update made available by the rst sentence (ignoring problems of modality)
is some party. The only set of individuals that can in any sense be said to be
specied as part of the context is some set as witness to the expression a party,
or in some circumstances its complement set.16 But even allowing that the
complement of a witness set might exceptionally form part of a specication
of context, this would not alleviate the problem that this is not suÆcient to
guarantee the unique individuation of the person successfully referred to by
the use of the pronoun she as part of the context relative to which it denes
an update. What is required here is a concept of inference dened over the
information provided by the utterance of the antecedent sentence { in this case
along the lines `If A is having a party, C will be there. A is having a party.
Therefore C will be there.' But the need to construct premises from which
the implication of the presence of C at the party is derived is driven by the
presence of the pronoun itself. The specic entities depicted in the situation are
in no sense available in the context within which the sentence is uttered. So the
concept of context is quite unlike the content-independent context presumed in
a semantics in terms of context-updates.17
16 Note the use of the complement set as the basis for establishing anaphor-antecedent rela-
tions in negative sentences such as (i), an observation noted by Corblin (1996), and developed
by Kibble (1997):
(i) They didn't invite a professor. She would have spoiled the party.
17 This is the problem of presupposition accommodation (see Beaver 1997 for an overview).
But notice that in this case accommodation must not have the eect of updating some party
context to include the designated individual. Formal constructive theories of context have
been proposed by van der Does (1996) and Rooth (1992, 1996).
Pronominal Anaphora: Semantic Problems 15
1.4.4 Quantication
The nal problem for the semantic account of anaphora is the use of quantied
expressions which enable an antecedent to be made available for construal of
a following pronoun that is not part of the interpretation of the quantied
expression itself. In (1.33) two situations are described disjunctively, and the
pronominal her then refers to some individual indicated in whichever of the two
situations holds:
(1.33) Either we can get a girl to read a poem, or we can get an older woman
to rattle o a passage from Yeats. If it takes her 10 minutes,...
In such cases, one might argue, a pair of situations is introduced as a context
of evaluation (as in Milward 1995), with no means of choosing between them.18
What is needed is a step of inference derivable from the two situations severally
described to a third situation containing someone who reads for whoever `us'
is. But no such individual is available in the situation corresponding to the
interpretation of the sentence. To side-step such problems, Milward suggests
a concept of meta-level disjunction, without any ability to distinguish between
the disjuncts. He posits an arbitrary name mechanism, the name in some sense
picking out either one of the two individuals, both being `in attention'. But this
is to directly incorporate proof-theoretic, and hence syntactic, constructs into
the semantic characterization. It is simply the rule of Disjunction Elimination
of natural deduction systems in a dierent guise (see Prawitz 1965, Lemmon
1965).
1.4.5 Syntactic Processes of Anaphora
Notwithstanding all these possible extensions to the semantic vocabulary, some
cases of pronominal anaphora remain intransigent. Anaphora may be integrated
into the syntactic system through the use of resumptive pronouns in both topic
and focus constructions and in relative clauses, and all such phenomena are
standardly taken to be unrelated to the central cases of anaphora. In certain
circumstances, a pronoun may function as a gap, with the pronoun appearing to
be subject to the very same subjacency constraints as syntactic gaps (Swedish
{ Maling and Zaenen 1981):
(1.34) *Lisa talar jag med pojken som kysst (henne) [Swedish]
Lisa talked I with the boy who kissed (her)
`As for Lisa, I talked with the boy who kissed her.'
18 These cases are highly problematic for DRT also, which imposes the requirement of intro-
ducing a new discourse referent for each indenite, thereby guaranteeing that there is not a
uniquely determined antecedent for the pronominal. Moreover, the embeddability conditions
for the conditional which follows require that all partial models in which the antecedent holds
must be extendable to that of the consequent. But the individual to be constructed in that
subsequent DRS cannot be identied as some girl or some woman as previously set up, be-
cause no such entity is introduced at the top level of the DRS constructed from the previous
sentence. The analogous problem applies to Dynamic Predicate Logic, for which they are
equally problematic.
16 Towards a Syntactic Model of Interpretation
18
A Preliminary Sketch 19
specication of some structure or formula by an independently accessible
representation.
As these ve assumptions demonstrate, the commitment is resolutely repre-
sentationalist, and the model will duly be dened in terms of how structured
representations are progressively established.
Having set out this statement of what we are going to be modelling, we must
swiftly add a codicil. What we shall not be modelling is how a parser succeeds
in making choices between dierent possible interpretations of a string or in
pragmatically establishing indirect inferences of various sorts. This, we assume,
depends on a broad range of factors, by no means exclusively structural, con-
strained by overarching considerations of relevance. What we are assuming in
putting such relevance-based considerations aside is that the modelling of nat-
ural language understanding presents two separate problems. The rst is one
of dening on a left-to-right basis how a structural representation of interpre-
tation for a string can grow { setting out, in other words, a formal architecture
for describing the parsing process. The second is to take such a framework and
articulate within it criteria which enable a parser to make choices that lead to
appropriate interpretations of a given string with the further pragmatic eects
that such choices give rise to. Our concern in this book is only with the rst of
these: we will articulate a formal model of interpretation growth for a sequence
of expressions, dening possible interpretations for strings of a language on a
left-to-right basis.
To give substance to the assumptions we have set out, a number of early de-
cisions have to be made. We have to choose some logical language for expressing
the representations which re
ect the content assigned to any given sentence. We
have to decide how to induce the structure, relative to which the rules of com-
bination for such formulae can be dened. And we have to be able to dene
the ways in which structures can be incomplete, with an appropriate process of
completion.
2.1 A Preliminary Sketch
The deductive system we need will be a composite one { dened in terms of
sub-modules and their interaction. We shall introduce it step by step. First
we shall sketch out the various sub-modules and the dynamics of the system
relatively informally, to give the more casual reader an overview. Then we shall
return to give more detailed denitions and exemplication.
The overall intuition we are modelling is that interpretation is built up from
left to right as each individual word is processed, following the combinatorial
properties of the words as specied by their logical type. Words are assumed
to project expressions in some logical language, and it is these that combine
together to result in a logical form corresponding to the interpretation of the
sentence. This much is familiar, like many categorial grammar systems in par-
ticular. However, we do not assume (as in Steedman 1996, Milward 1995) that
20 The General Framework
(a)
John
(b)
John
xy [Admire(x)(y )]
(c)
John
?t
John ?(e ! t)
?t
John ?(e ! t)
xy [Admire(x)(y )] ?e
?t
John ?(e ! t)
Mary
Then, with the words that follow, we get rst a pair of nodes, the rst anno-
tated by a formula projected by John, the second decorated by the requirement
?(e ! t):
?t
xy [Admire(x)(y )] ?e
At this point, the string ends; so, according to our goal-directed enterprise of
establishing some propositional formula, the string has supposedly presented
enough information to establish a proposition of type t. With one node that
requires an annotation and an annotated node requiring a location in the tree,
the latter can be taken to supply the annotation for the rst:
24 The General Framework
?t
John ?(e ! t)
and the tree is duly updated to yield the propositional formula Admire(Mary)
(John) { in other words, the same conguration that results from (2.1):
Admire(Mary )(John)
assumed that the point of departure in any account of parsing is the radical
underspecication of discrete alternative interpretations of a string, and that
the prime focus is to articulate resolution strategies that identify a denitive
structural description of some syntactic structure (see e.g. Marcus 1980 and
the following tradition of syntactic parsing and the characterization of semantic
underspecication explored in e.g. Pinkal 1995). However, we have chosen the
former option, and have set aside the problem of disambiguation as a secondary
task.
One of the principal reasons for setting the phenomenon of disambiguation
aside is the commitment to dene content of a linguistic expression as the contri-
bution it makes to building up an interpretation (see the discussion of anaphora
in chapter 1). An important distinction which we believe needs to be preserved
in such an account is that some expressions may make two or more distinct
contributions to the interpretation and so are ambiguous, while others have a
only a single contribution which may, however, not be fully determined. In
both cases the supercial phenomenon is one of availability of choice and hence
parsing uncertainty, but this masks the fact that the underlying phenomena in
the two kinds of case are dierent. This distinction is familiar in the case of
lexical content, where it is well established that a distinction needs to be made
between the English word bank (with its two quite discrete interpretations) and
the vacillation in interpretation of a pronoun relative to context, despite the fact
that, from a parsing perspective, it may be equally impossible to assign an in-
terpretation to either, in default of an appropriate context. Any account which
re
ects parsing uncertainty directly without dierentiating these cases has no
basis for distinguishing idiosyncratic lexical ambiguity from systematic cross-
linguistic underspecication of interpretation.2 We believe the same distinction
applies at the structural level. While the initial constituent in a left-dislocation
structure may underspecify its interpretation within some unfolding structure,
this is dierent in kind from the uncertainty confronting a parser at the on-
set of a string whose rst constituent is a noun phrase, where the uncertainty
is a consequence of the availability of choice between more than one discrete
structure.
This view is buttressed by work in pragmatics, where there is evidence that
disambiguating a structure and enriching weak specications of content are pro-
cesses subject to the same psychological constraints, but that these none-the-less
constitute quite dierent phenomena (see Sperber and Wilson 1995). Moreover,
the problem of establishing a particular interpretation in any given context is an
essentially cognitive one { according to Relevance Theory concerning interaction
between minimizing cognitive eort for establishing a suÆciency of inferential
eects. Whatever the precise form of these constraints, the problem of dening
them is secondary to the task of dening the process of tree growth involved in
the interpretation process and the form of input that expressions contribute to
the growth of such structures.
2 See the discussion in chapter 1. It should be noted that some accounts of semantic
underspecicity adopt the stance that there is no distinction in principle between these two
cases: see Poesio 1996, van Deemter 1996, Pinkal 1995.
The Data Structures of the Parsing Model 27
Accordingly, we have elected to dene parsing of a sentence as the unfold-
ing of tree descriptions for all possible interpretations. In so doing, we choose
to distinguish not merely between lexical underspecication and lexical ambi-
guity, but also between structural underspecication and structural ambiguity.
And in setting out the landscape of possible construction steps, the framework
articulates all possible transitions for a given sequence of words that yield a
well-formed logical structure, presuming on a pragmatic theory such as Rele-
vance Theory (see Sperber and Wilson 1995) to provide the criteria for choosing
between them. In the process we develop an account of long-distance depen-
dency in terms of initial structural underspecication and subsequent resolution
of uncertainty. We return to the question of whether this system is too liberal
at several points during the book, in particular in chapter 6.
The system we describe is modular, and uses the terminology of Labelled
Deductive Systems (LDS), of Gabbay (1996). Labelled deductive systems are
systems in which deduction is dened for Label{Formula pairs. The components
represented by the Label and the F ormula are independently dened systems,
and the Labelled Deductive System in which they are embedded provides the
logic of their combination. In our system, the formulae represent sub-terms
of logical forms, and the labels represent specications that drive their com-
bination. The operations that combine formulae are set within a goal-directed
enterprise of constructing an overall logical form. In this construction process, a
tree structure is progressively introduced and decorated with labelled formulae.
The process is driven by instructions projected by the sequence of words of a
natural language string.
We introduce the system in three stages, represented diagrammatically in
gures 2.4{2.6. Initially we present the basic structures and a vocabulary for de-
scribing them (in this chapter). This characterization will necessarily be rather
abstract, since every phenomenon explained in subsequent chapters involves the
characterization of how partial structures grow; and to see this, we have rst to
set in place the vocabulary for describing individual partial trees. We then de-
ne the dynamics involved in building up individual tree structures (chapter 3) {
covering simple clauses, complement clauses and long-distance dependency. We
then extend the system to account for relative clauses (chapter 4), introducing
a concept of linked tree structures.
2.2 The Data Structures of the Parsing Model
In our parsing system, so-called Formulae are used to express the denotational
content of individual terms, phrases and sentences. Arguably, the construction
of these Formulae is the central goal of the parsing process, and in our system
they will take the form of terms of a familiar typed lambda calculus. The mode
of combination of these formulae is determined by their logical type, and these
28 The General Framework
In this chapter the two types of Data Structures are introduced that are used by our
(theoretical) parser: rstly Basic Tree Structures, the representations of Logical
forms (denition 2)and secondly Partial Tree Structures (denition 5), the
objects the parser constructs on its way to full Basic Tree Structures.
Basic Tree Tree Structures consist of binary trees the nodes of which are decorated
by Declarative Units (section 2.2.1). The following structure represents the meaning
of John admires Mary:
h: : : ti : Admire(Mary)(John)
h: : : ei : John h: : : e ! ti : y[Admire(Mary)(y)]
h: : : ti : Admire(Mary)(John)
h: : : ei : John
h: : : ! ti : y[Admire(Mary)(y)]
h: : : e ! (e ! t)i : xy[Admire(x)(y)
The tree structure is addressed by tree The tree decorations are addressed by La-
modalities (section 2.2.2), tree node bel and Formula predicates, for instance,
identiers (section 2.2.6.2), and by the the Label predicates T y (section 2.2.1)
(sub)language ADR (denition 7). and T n (section 2.2.6.2) and F o, the For-
mula predicate (section 2.2.1).
Partial Tree Structures consist of structures that can develop to Basic Tree Struc-
tures. The nodes of Partial Tree Structures are decorated by Partial Declarative
Units and Requirements (section 2.2.5)
[ T y(t); h#1 i(T y(e ! t))]
The Partial Tree Structures are addressed The Tree Decorations consist of pairs of
like Basic Tree Structure, but there is a (partial) declarative unit, to the left of
an additional distinction between Exter- the bullet and a set of requirements to
nal and Internal Tree Modalities (section the right of the bullet (denition 5).
2.2.5.1).
types we will display as Labels to the lambda term which is the corresponding
Formula.3
The simplest mode of combination might be represented as the following
application of what is in eect a step of Modus Ponens on the side of the Type
label and function application on the Formula side:
e : John e ! t : Smile
t : Smile(John)
Lexical items of the natural language project, among other things, the requisite
labelled formulae to act as premisses. These are not the words themselves
3 The types, taken from type theory as in the Montague grammar tradition, are expressed
as in type-logical grammars as propositional formulae in a conditional logic: the semantic
type he; ti corresponds to the formula e ! t, he; he; tii as (e ! (e ! t)), etc. See Morrill 1994,
Moortgat 1997, Oehrle 1995.
30 The General Framework
but concepts which the word is taken to express. Though the types provide
the major label driving the combinatorial process, other labels may provide
additional control, including gender specications, scope restrictions, labels for
discrete sentence types, etc. The unifying characteristic of these labels is that
they serve as controls on the process of setting out and combining the formulae
projected by the lexical expressions. Initially we shall assume that the only mode
of combination consists of the twinned operations of functional application and
type deduction, as indicated in the above step of labelled Modus Ponens, but,
in chapter 4 we will dene a mode of combination which resembles an update
action to represent the incorporation of information of a relative clause into the
head.
We will introduce in tandem our basic data structures, decorated partial
trees and the language DU to talk about them (DU for Declarative Unit). In
this section we will not be wholly formal but will introduce our structures by
highlighting three typical features of the account which together bring out the
The Data Structures of the Parsing Model 31
form of the descriptions we shall give and the dynamics of how such descriptions
grow:
the atomic propositions of the language DU
the tree modalities the language DU uses to describe relations within a
tree structure
the representation of requirements in DU to dictate directions in which a
tree is to develop
2.2.1 Atomic Formulae
In the following paragraphs, we rst dene the Type and Formula predicates
and introduce the propositions which serve as decorations of the tree structure.
We shall dene Labels and Formulae initially by means of monadic predicates
in a predicate-argument notation, but in the nal chapter we shall provide a
more general denition in terms of feature structures. These propositions we
then embed in a tree description language which will be adequate for describing
logical forms in decorated tree representation.
As we have so far sketched, the aim of our parser is to construct decorated
tree structures. The decorations on the nodes of such trees consist of declarative
units. Declarative units are pairs consisting of a sequence of labels followed by
a (content) formula:
h|l1 ; :{z: : ln}i : |{z}
.
Labels F ormula
The general methodology behind such labelled formulae is that the formula is the
primary term, with labels providing additional information about that formula
{ its potential for combination with other formulae, its semantic evaluation, its
position in a structure, and so on (see Gabbay 1996). We have seen examples of
formulae in the denotations John, Mary and xy[Admire(x)(y)]. The types
e, t and e ! t from these examples are instances of labels. For instance, the
decoration of the root node of gure 2.2 is the declarative unit
h: : : ; ti : Admire(Mary)(John):
It is the language DU which expresses facts about these declarative units. The
atomic propositions of the language have the form P (a), where P is a monadic
predicate letter and a is some appropriate term. For each Label category of the
declarative units we will have a predicate Lai, and for the Formula category we
have the predicate F o. The above declarative unit, for instance, supports the
facts F o(Admire(Mary)(John)) and T y(t), which we will display as
h: : : ti : Admire(Mary)(John) j= T y(t) ^ F o(Admire(Mary)(John)).
32 The General Framework
Here du j= Lai(li ) means that the formula Lai(li ) holds at, or is satised by,
declarative unit du.
The parser we are constructing takes a natural language string to produce
a logical term in some formal language. This term we represent in the form
of a decorated binary tree. So our data structures will be decorated trees. To
describe these structures, however, we need a (second) formal language, so
we introduce a language to talk about data structures which themselves are
formulae in some language.
When a proposition, say P (a), is true of an object hl1 : : : lni : we will use
the notation
hl1 : : : ln i : j= P (a):
That is, we separate the object on the left from the proposition on the right by
the symbol `j='. If P (a) is true on object hl1 : : : lni : , we will also say that
hl1 : : : ln i : satises P (a) or that P (a) holds on hl1 : : : ln i : .
So, as displayed above, the declarative unit h: : : ti : Admire(Mary)(John)
satises both the proposition T y(t) expressing that the unit has the Type label t
and the proposition F o(Admire(Mary)(John)) stating that the content formula
of the unit is Admire(Mary)(John). This is an instance of a general pattern.
A declarative unit hl1 ; : : : lni : can be completely represented by a description,
the nite set of atomic propositions satised by that unit:
fLa1(l1 ); : : : ; Lan (ln ); F o( )g:
So in this description, both labels and formulae are represented by monadic
predicates. A partial declarative unit, an object naturally arising in the course
of a parse, is now merely a subset of a description of a declarative unit. The
descriptions of declarative units determine the atomic vocabulary of the lan-
guage DU . So, this language has monadic predicates La1; : : : Lan; F o, standing
for n Label dimensions and a Formula dimension and individual constants from
DLa1 ; : : : DLa ; DF o respectively, denoting values on these dimensions. The sub-
sequent paragraphs will discuss the Type label T y with values in the domain
n
of the form
T y(U ! (e ! t))
as annotations of tree nodes. Notice that U ! (e ! t) 62 DT y : this schematic
type can end up as types e ! (e ! t); t ! (e ! t), (e ! t) ! (e ! t), etc. by
updating, but is not itself a type (see Marten 1999).
The Formula Predicate F o
The Formula predicate F o has as values elements in the set DF o consisting of
terms of a typed lambda calculus. These terms express the content assigned
to a natural language string, and, arguably, the construction of a value of the
F o predicate, given a natural language string, is the whole point of the parsing
process. The Formula predicate and the Type predicate work in tandem in the
basic tree conguration
fF o( ()); T y(Y )g
{ t1; t2 2 DLa [ DF o [ V AR [ MV ,
{ # an element of f#0; #1; "0; "1; #; "; #; "g
i
T ; n j= ;
if
6= ?
=>
= Lai (a) and li = a
= F o() and =
= Eq(t1 ; t2 ) and t1 = t2
= ^ and T ; n j= and T ; n j=
= _ and T ; n j= or T ; n j=
= 9x and there is a t 2 D : T ; n j= [t=x]
10 The re
exive transitive closure of the relation is the set of all ordered pairs hn; n0 i such
that there is a natural number k, possibly 0, such
n : : : n0 :
| {z }
k times
11 The Basic Tree Structures T interpret only the `static' logical connectives and operators:
`_', `^', and `9x', and `h#i'. The remainder of the standard connectives and operators, i.e.
`:', !', `8x' and `[#]' will be discussed shortly.
40 The General Framework
if i 2 f0; 1g and
= h#i i , and 9n0 2 T rDom : n i n0 and T ; n0 j= ,
= h"i i , and 9n0 2 T rDom : n0 i n and T ; n0 j= ,
= h#i , and 9n0 2 T rDom : n n0 and T ; n0 j= ,
= h"i , and 9n0 2 T rDom : n0 n and T ; n0 j= ,
= h# i , and 9n0 2 T rDom : n n0 and T ; n0 j= ,
= h" i , and 9n0 2 T rDom : n0 n and T ; n0 j= .
Notice that no decorated node n satises ?; > holds on every node, and the
atomic formulae hold if the declarative unit annotating the node has the right
label and formula values. The Boolean connectives and the modal operators
have their standard interpretation.
The meta-variables can be distinguished from the proper values by the fact
that only for a proper value li we have the satisfaction of 9xLai (x). 9x holds
at a node if [t=x] holds there for some t 2 D = DLa [ : : : DLa [ DF o. We
want the existential quantier to be able to express that some label or feature
i n
2.2.5 Requirements
At a particular state in the parsing process, a (partial) tree will have been
constructed, the nodes of which are decorated on the one hand by (partial)
declarative units and, on the other hand, by requirements to be satised. The
root node, for instance, will have a requirement that it be annotated by a lambda
term of type t, a sub-node of the root that it be annotated by a term of type
e, and so on. Each of these requirements is a proposition in the DU language.
This use of requirements on the development of a tree node has some
resemblance to the familiar concept of subcategorization, as a node decorated
by the labelled formula
h: : : ; (e ! (e ! t))i : Admire
within a tree may have a mother node which is decorated with a requirement
h#0 iT y(e) (i.e., a requirement for an internal argument for Admire). However,
the concept of requirement is much more general than subcategorization state-
ments: all nodes are introduced with requirements. Requirements form an
essential feature of the tree { they determine the set of successful extensions
of a given (partial) tree: namely, those in which all requirements are satised.
Consequently, our basic data structures are tree structures, the nodes of which
consist of (partial) declarative units paired with nite sets of requirements.
A requirement can be any formula in the DU language under development.
A tree node will now be decorated by a pair consisting of a partial declarative
unit du and a nite set of requirements R:
[du R]
Here the bullet separates the body of the declarative unit from the requirements.
In the language DU which describes these nodes, we will introduce a question
mark to represent requirements, that is:
[du R] j=? i 2 R.
Below is an example of a partially decorated tree T that occurs in the construc-
tion of John admires Mary, (2.1):
[ T y(t); h#1 i(T y(e ! t))]
than l. The specic content of this notion depends on the specic Label
i
U La a: i
La (i.e. elements l such that l La l implies that l = l ) then represent
normal forms under this rewrite relation. For declarative units du; du0 we
i i
can now set du du0 if for all 0 < i < n + 1 : li La li0 and F o 0 .17
i
nal complete tree structure, then we cannot let T ; n j= [#] only to mean that
holds at all nodes in T dominated by n.
The connectives and operators with universal force are necessary when we
want to express invariances, universal principles which our partial trees should
satisfy. These principles pertain to the axiomatization of the tree structures
themselves, but also to invariances pertaining to Label and Formula values,
their interactions and co-occurrences. With respect to the latter case, we will
have to introduce axioms regulating the behaviour of the F o and T y predicates
on the trees. For instance, a node may be annotated by at most one type. This
requires a principle of the form:
8x8y((T y(x) ^ T y(y)) ! Eq(x; y)):
Furthermore, the T y values at the daughters of some node (and also the F o
values) have to be related to the values on those predicates at the node itself
8x8y((h#0 iT y(x) ^ h#1 iT y(x ! y)) ! T y(y)):
The structure of the tree is regulated by LOFT principles of the form
{ h#0i ! :h#1i
{ h#0i ! h#i
{ h#0i ! [#0]
The rst of these principles states that an argument daughter is never a function
daughter. The second expresses that an argument daughter is always a domi-
nated node, and the third that there is essentially only one argument daughter.
Denition 6 (Satisfaction II) In terms of the relation of extension between
partial trees, we now reinterpret the evaluation of atomic formulae and give
the interpretation of the missing universal connectives and operators. So we
continue the denition of satisfaction. Node n = [hl1; : : : ; lni : R] of Partial
Tree Structure T satises DU formula , with the notation
T ; n j= ;
if
= Lai (a) and a La li
= F o() and F o
i
b c
which can be extended both downwards and upwards, from the nal structure
["]?
a; [#]?
b; [#]? c; [#]?
which can no longer be developed to a structure in which the present frontier
nodes become internal ones.
2.2.5.1 Internal and External Modalities
The denition of the notation T ; n j= for the static Boolean and modal oper-
ators is identical to the notion T ; n j= as given by denition 3. But on these
partial structures there are in eect kinds of each underspecied modality. Sup-
pose, for instance,
T ; n j= h# i:
Thus, by denition, there is some n0 such that n n0 and T ; n0 j= . This
n , however, may be otherwise wholly unconnected to the daughters and grand-
0
daughters of n. We know that the partial tree can be extended to one where
n0 is dominated by n, but that is all. Let's call h# i the internal version of the
dominance modality.
On a partial tree we can, however, also interpret a `real' relation of dominance.
Let us introduce the symbols # and # with the meaning
T ; n j=# i n i n0 for i 2 f0; 1; #g and T ; n0 j= ,
T ; n j=# i there is a nite sequence n i1 : : : i n0 for
ij 2 f0; 1; #; g and T ; n0 j= .
n
We will call these the external variants of the modalities h#i and h#i respec-
tively. Figure 2.9 shows a situation where the two variants of h#i dier in their
interpretation. In the Partial Tree T of that gure, evaluation at the top node
gives T ; [ T y(t)] 6j= h#iF o(John) but T ; [ T y(t)] j= # F o(John). Similarly,
T ; [ T y(t)] j= # F o(xy[Admire(x)(y)]) but not necessarily T ; [ T y(t)] j=
h# iF o(xy[Admire(x)(y)]). However, T ; [ T y(t)] j= h# iF o(Mary) and
The Data Structures of the Parsing Model 49
T ; [ T y(t)] j= # F o(Mary) By denition of a partial tree structure as extend-
able to a basic tree structure, we know that T ; n j= h#i implies that every
extension T 0 of T (T T 0) has an extension T 00 such that T 00 ; n00 j=# : the
internal variant can always be extended to the external one.
[h: : : i : T y(t)]
language is consistent, i.e. has some model, this model is actually created from
the syntactic entities of the language itself (consistent sets of formulae become
domain elements, names become the objects they stand for, etc.).
Every state of the parsing process is a tree description : that is, every state is
a consistent family of coherent sets. In general, this consistency is not something
that has to be shown or concluded; it is something that is preserved by the
transition rules of our system. Moreover, if the natural language string is a
grammatical sentence, then the nal description is isomorphic to a complete
binary tree without requirements { that is, a term in the logical representation
language serving as the interpretation of the natural language string.
In this section we will introduce rst node descriptions and then tree descrip-
tions as collections of descriptions connected by relations dened on the basis of
their `internal' structure. That is, here we deal with the `states' of the transition
system. The next chapter will then introduce the `transitions' between these
states that are induced by parsing a natural language string.
2.2.6.1 Descriptions of Nodes
A declarative unit plus requirements [du R] can be completely described by a
nite set of DU formulae: the union of the set of atomic DU formulae satised
by du and a set of requirements satised by that unit, i.e. the DU formulae ?
such that 2 R. For instance,
[h: : : ; ei : John h"iT y(t)]
can be represented by
f: : : ; T y(e); F o(John); ?h"iT y(t)g.
This means that the nodes of a decorated tree can be represented by a nite
family of nite sets of DU formulae.
Let us call a nite set of atomic DU formulae and requirements representing
a declarative unit plus requirements a DU set. Not every nite set of atomic
DU formulae and requirements is a DU set. For instance, if a set contains both
T y(e) and T y(t), then it cannot represent a declarative unit: a nite set of DU
formulae must be consistent with all information about the labels; in particular,
it must satisfy specic co-occurrence restrictions. The co-occurrence of two
values on the type label is excluded; another instance is the interdependence
of the content formulae and the type labels. On the other hand, if a set of
DU formulae is `consistent' with the principles of declarative units, then it may
contain more than atomic DU formula and requirements. So also then, it will
not (strictly) represent a declarative unit plus requirements. A description of
the declarative unit du may contain facts that can be computed or derived from
the elements of the DU set corresponding with du. That is, a description of du
may contain Boolean combinations of atomic DU formulae and requirements.
But there is more. A node description may also contain formulae headed by
a tree modality. Thus, a node description may represent a node within a
tree environment. For instance, the set Z = fF o(); T y (X ); ?h#1 iT y (Y );
The Data Structures of the Parsing Model 51
h#0 iF o( )g describes a node in the context of an argument daughter which
satises F o( ):
[h: : : ; X i : h#1 iT y(Y )]
[h: : : i : ] [ T y(Y )]
A tree description containing only the set Z would not be coherent. It should
contain another description which is consistent with being an argument daugh-
ter. Consistency of descriptions must now take into account the LOFT principles
which determine the Logic of Finite Trees. Thus, if a description states that
holds at an argument daughter (h#0 i) and that holds at an argument daugh-
ter (h#0i ), then the description of that argument daughter must be consistent
with ^ , for it is a LOFT principle that there is at most one such daughter.
Now we can give an informal denition of a node description, as a nite set of
formulae that can be satised by some declarative unit in a partial tree structure
such that no formula is both established and required.
2.2.6.2 Descriptions of Trees
In order to code the structure of a decorated partial tree T ; n in descriptions,
we will introduce a monadic tree node predicate T n and a pointer }. A formula
of the form T n(a) will be called a tree node identier. Node descriptions can
be connected by a tree relation through the presence in them of tree node
formulae in combination with the tree modalities of the DU language; this gives
a description of T . The pointer } is an element of exactly one node description,
and this node describes n. For instance, the tree T ; [h: : : ei : John], where T is
the tree
h: : : ti : Admire(Mary)(John)
following picture:
fT n(a); : : : g
fT n(a0); : : : g fT n(a1); : : : g
fT n(a10); : : : g fT n(a11); : : : g
In a sequence a0, 0 represents the argument daughter of a; in a1, 1 represents
the function daughter of a. Thus a10 indicates the argument daughter of a node
that is the function daughter of a node with address a.
A description containing T n(0) represents a top node of a tree that will
remain a top node. When trees are constructed in the parsing process, in
general it is not known whether a description that starts o as a top node will
remain so (and thus be the root node of the eventual tree). This is why a new
node description is invariably introduced containing a formula T n(a), where a
is a constant not yet occurring in the construction. In interaction with the tree
modalities, various constellations are expressible. So, given the formula T n(a)
expressing the location of a node description in the tree under construction, we
can set
T n(a0) $ h"0 iT n(a),
T n(0a) $ h#0 iT n(a),
T n(a1) $ h"1 iT n(a),
T n(1a) $ h#1 iT n(a),
and we can x the root node of a tree as follows:
T n(0) $ ["]?:
The tree node formula T n(0) holds at a node if it is a top node and remains
so throughout all developments. Note the use of the `falsum' { `At every node
above the current one, ? holds'. As ? is satised by no node at all (denition
3), this means that there are no nodes above the current one.
Here are some examples of addressing in the node descriptions:
fT n(a); : : : g. This set need not be a top node, but it could be.
fT n(a); ["]?; : : : g. This must be a top node, so
fT n(a1); ["]?; : : : g is inconsistent (after all, T n(a1) $ h"1 iT n(a)). When
all words of a string have been processed, the the current top node can be
declared the root node of the tree. That is, the top node can be closed o
The Data Structures of the Parsing Model 53
by adding ["]?. Then, by denition, we have T n(0), and this is an update
of the proposition T n(a) previously present in the same description (for
we have a T n 0).
fT n(a); [#]?; : : : g. This must be a terminal node. Bottom nodes are
closed o downwards { that is, declared terminal nodes by the words of a
natural language string. Again, note the use of the falsum.
f: : : h" iT n(a); : : : ; ?9xT n(x)g. In this node description, the requirement
can be fullled only when the description is connected to the description
with address T n(a) by a sequence of 0; 1 steps: only then is there
a t 2 DT n such that the description satises T n(t). This means that
we can combine a structurally underspecied address, h"iT n(a), with a
requirement such that the node can only rid itself of all requirements by
becoming structurally fully specied. It is this characterization, using the
Kleene star operator, which expresses the weak tree relation between a
left-dislocated constituent and the node at which it is introduced.
Now, by representing a decorated tree as a family of node descriptions struc-
tured by the T n predicate and the tree modalities, we create the possibility
for computational and deduction rules which take the content of more than
one description into account. First, we introduce the notion of the language of
addresses as a sub-language of DU .
Denition 7 (Addressing of Node Descriptions) The Address Language
ADR is the smallest set such that
>; ?; T n(t) 2 ADR for t 2 DT n
if 2 ADR and h#i, [#] are tree modalities, then h#i; [#] 2 ADR.
Now we use the language ADR to dene tree relations between node descrip-
tions.20
Denition 8 (Tree Relations between Node Descriptions) If ND and
ND0 are node descriptions and # 2 f#0; #1 ; #; #g, then
ND # ND0
if there is a 2 ADR such that either 2 ND0 and h#i 2 ND, or 2 ND
and h# 1 i 2 ND0 . We set ADR(ND) = ADR \ ND. This set gives all
formulae from ND that are involved in the determination of its location in the
tree.
On the basis of this system of relating node descriptions by the tree relations we
can formulate what it is to be a tree description. To describe a basic or partial
tree structure T , we have to give a description for every node; and if we want
to interpret the relations we have dened above as the real tree relations, then
20 For fully explicit denitions, see denitions 22{24 in chapter 9.
54 The General Framework
55
56 The Dynamics of Tree Building
2 Recall that h#0 i is interpreted along the argument daughter relation, h#1 i along the func-
tion daughter relation.
58 The Dynamics of Tree Building
Tree Description V
fT n(a); h#0 iT y(e); ?h#1 iT y(e ! t); ?T y(t)g
fF o(Admire)g
f fT n(a); ?T y(t); h#0iT y(e); ?h#1 iT y(e ! t)g
fh"0 iT n(a); F o(John); T y(e); [#]?g
fh"1 iT n(a); ?T y(e ! t); ?h#0 iT y(e); }g
fh"1 ih"1 iT n(a); F o(Admire); T y(e ! (e !)); [#]?g g
In displaying the trees, we tend to mimic the relative order of words in the lan-
guage, but nothing turns on this. Word order is not re
ected in the structure
itself, only in the order in which lexical actions take place.
The Parsing Process { A Sketch 61
Step 7: Prediction
Prediction can develop the requirement for an object h#0 iT y(e) of admire into
an argument daughter of the VP node with requirement ?T y(e).
It should be noted that these Scanning and Prediction steps, induced by the
lexicon, can be mimicked by rule applications: by a repetition of Introduction
and Prediction now with respect to the ?T y(e ! t) requirement at the predicate
node. First we apply Introduction to
f?T y(t); ?h#1 iT y(e ! t)g
Then we unfold the rst modal requirement at the predicate node by Prediction,
which results in the same tree structure as description VI. So there are choices
involved in what to assign to the lexicon and to general principles. And we will
see that individual languages dier in the way they divide the labour between
the lexicon and general principles.
Step 8: Processing the pronoun
The object him is then scanned in the same way as the word John. The action
connected to this pronoun is of the form:
IF ?T y(e);
THEN add F o(U); T y(e); ?9xF o(x); ?h"iT y(e ! t);
ELSE ABORT.
The variable that this pronoun introduces is accompanied by the requirement
for 9xF o(x), which is satised only by a full formula value (not another meta-
variable), and the accusative case projects a requirement for h"iT y(e ! t) which
is fullled if the node is a daughter of a predicate-type node.4 Executing this
action results in the description:
4 We ignore here the problem that prepositions also require this accusative form.
62 The Dynamics of Tree Building
Tree Description I
fT n(a); ?T y(t); }g
Step 2: Adjunction of a dominated node
Now we allow the introduction of an unxed node, as follows:
Tree Description II0
fT n(a)g
f?T y(e); }g
f fT n(a); ?T y(t)g
fh" iT n(a); ?T y(e); ?9xT n(x); }g g
This describes a truly partial tree, because it contains an connection that
cannot be analysed in a sequence of 0, 1 or # connections. This move is an
5 Shortly, we shall return to the issue of whether these two modes of interpretation should
be distinguished at this early point in the interpretation process.
The Parsing Process { A Sketch 65
example of the rule of *Adjunction. We have created a location for an argument
in the term under construction, without specifying where this argument is to be
located within the sub-term structure. This underspecication with respect to
its location is witnessed by the requirement ?9xT n(x), which is only satised
once the node has a fully specied location { that is, for T n(x), where xf0; 1g.
This extension of tree description I, the starting point, is an alternative to
the development shown in tree description II of the previous example. The
alternatives present for our parser are determined by the natural language under
observation
tree description I
tree description II or tree description II0 or : : : or : : :
where tree description II results from description I by an application of the In-
troduction rule, whereas tree description II0 results from the same description I
by applying the *Adjunction Rule.
Step 3: Scanning
Now the word Bill maps this structure onto one where the unxed node receives
F o(Bill) and T y(e) and the pointer has returned to the top.
Tree Description III
fT n(a); }g
Tree Description V
fT n(a)g
fh" iT n(a); fF o(John)g
F o(Bill)g
fF o(Admire)g f?T y(e); }g
f fT n(a); ?T y(t)g
fh" iT n(a); F o(Bill); T y(e); ?9xT n(x); [#]?g
fh"0 iT n(a); F o(John); T y(e); [#]?g
fh"1 iT n(a); ?T y(e ! t)g
fh"1 ih"1 iT n(a); F o(Admire); T y(e ! (e ! t)); [#]?g
fh"0 ih"1 iT n(a); ?T y(e); }g g
Now the string is nished, but the structure still has a requirement outstanding
at a terminal node (the last node description). On the other hand, there is
surplus material in the unxed node (the second node of the description above).
Now, merging the unxed node with the pointed terminal node gives the nal
structure. In a tree description, merging of two nodes simply comes down to
taking the union of the corresponding descriptions:
fT n(a)g
fF o(John)g
In other words, noun phrases may either occur with an (as yet) completely
unspecied role in the developing logical form, or they may occur at positions
where their (combinatorial) role is strictly circumscribed.
We will give an example involving one level of subordination to show that
even the structure projected from verbs must be taken to be initially unxed
with respect to the ultimate root:
(3.5) Hiroto ga ringo o tabeta to itta
HirotoNOM appleACC ate COMP said
either `Tom/Dick/Hiroto/Akiko ... said that Hiroto ate the apple.'
or `Hiroto said that Tom/Dick/Hiroto/Akiko ... ate the apple.'
The rst step is common to all parse sequences: we start again from the Axiom,
but this time we x from the beginning that this will be the root node of the tree.
Tree Description I
fT n(0); ?T y(t); }g
The subsequent step is then an execution of the actions associated with the NP
Hiroto. This results in the following tree:
fT n(0)g
fT n(0)g
fF o(Hiroto)g f?T y(e); }g
fF o(P (; P ))g
fF o(U)g fF o(Ringo)g
f fT n(0); ?T y(t)g
fh"iT n(0); F o(Hiroto); T y(e); ?h"0 iT y(t); ?9xT n(x)g
fh"iT n(0); ?T y(e); ?9xT n(x); }g
fh"1ih" iT n(0); T y(cn ! e); F o(P (; P ))g
fh"0ih" iT n(0); ?T y(cn)g
fh"0ih"0 ih" iT n(0); T y(e); F o(U)g
fh"1ih"0 ih" iT n(0); T y(e ! cn); F o(Ringo)g g
The building of this local sub-tree with root node requiring T y(e) is induced by
actions associated with the common noun itself, because Japanese is a language
in which nouns need no morphological determiner. The resulting structure is
nevertheless, in default of any quantier, an existentially quantied expression.
As with the rst noun phrase sequence, the case suÆx o adds a requirement
to the root node of this locally constructed tree, here 11h"0iT y(e ! t), and the
case marker then returns the pointer to the root node.
fT n(0)g
F o(Hiroto) f?h"0 iT y(e ! t); }g
fF o(P (; P ))g
fF o(U)g fF o(Ringo)g
11 Here we ignore the dependency predicate and the selection of one of its arguments to
x the scope of the quantied term, though it should be noted that this sequence of unxed
nodes provides a good base for modelling the observation (Hoji 1986) that in Japanese scopal
choices are very largely driven by linear order, indenites apart (see chapter 7).
72 The Dynamics of Tree Building
fT n(0)g
fF o(Hiroto)g fF o(; x; Ringo(x))g f?T y(t); }g
fF o(V)g fh"1 ih" iT n(0)g
fF o(W)g fF o(T abeta)g
The Parsing Process { A Sketch 73
f fT n(0); ?T y(t)g
fh"iT n(0); F o(Hiroto); T y(e); ?h"0 iT y(t); ?9xT n(x)g
fh"iT n(0); F o(; x; Ringo(x)); T y(e); ?h"0iT y(e ! t)g
fh"iT n(0); ?T y(t); ?9xT n(x); }g
fh"1ih" iT n(0); T y(e ! t)g
fh"0ih" iT n(0); T y(e); F o(V); ?9xF o(x)g
fh"0ih"1 ih" iT n(0); T y(e); F o(W); ?9yF o(y)g
fh"1ih"1 ih" iT n(0); T y(e ! (e ! t)); F o(T abeta)g g
(In this tree display, we assume for simplicity that steps of Completion and
Elimination have taken place fully annotating the sub-trees projected by the
words Hiroto and ringo.) At this point, enough material has been collected
to create a complete decorated binary tree, i.e. a logical form. The meta-
variables which function as arguments for the predicate tabeta can be assigned
a value exactly as would a meta-variable projected from a morphological pro-
noun (as in the rst example). In particular, we can merge the nodes dec-
orated by F o(Hiroto) and F o(; x; Ringo(x)) with those decorated by F o(V)
and F o(W) respectively. We might also allow the nodes decorated by ?T y(t)
to be merged, yielding a fully specied structure with a root annotated by
F o(T abeta(; x; Ringo(x))(Hiroto)) (= `Hiroto ate an apple'). But, as the string
is not yet nished, this last step may well turn out to be premature.
The suÆx to now adds to the root node of this sub-structure a requirement
for a mother of type e ! t (i.e. its position in the eventual tree must be that
projected by a subordinate clause), and returns the pointer to the root. Now
the verb itta (`said') is dened as building a sequence of nodes in like manner
to tabeta, except that the argument daughter of its predicate node is required
to be of ?T y(t):
IF f?T y(t)g
THEN make(h#i); put(9xT n(x));
make(h#0 i); go(h#0 i); put(F o(X); T y (e); ?9xF o(x)); go(h"i);
make(h#1 i); go(h#1 i); put(?T y (e ! t));
make(h#0 i); go(h#0 i); put(F o(Y); T y (t); ?9xF o(x)); go(h"i);
make(h#1 i); go(h#1 i); put(F o(Itta); T y (t ! (e ! t)));
gofirst"(?T y (t))
ELSE ABORT
74 The Dynamics of Tree Building
fT n(0); ?T y(t)g
f fT n(0); ?T y(t)g
fh" iT n(0); F o(Hiroto); T y(e); ?h"0iT y(t); ?9xT n(x)g
fh" iT n(0); F o(; x; Ringo(x)); T y(e); ?h"0iT y(e ! t)g
fh" iT n(0); ?T y(t); ?9xT n(x)g
fh"1 ih" iT n(0); T y(e ! t)g
fh"0 ih" iT n(0); T y(e); F o(V); ?9xF o(x)g
fh"0 ih"1 ih" iT n(0); T y(e); F o(W); ?9yF o(y)g
fh"1 ih"1 ih" iT n(0); T y(e ! (e ! t)); F o(T abeta)g
fh" iT n(0); ?T y(t); ?9xT n(x); ?h"0iT y(e ! t)g
fh" iT n(0); ?T y(t); ?9xT n(x); }g
fh"1 ih" iT n(0); T y(e ! t)g
fh"0 ih" iT n(0); T y(e); F o(X); ?9xF o(x)g
fh"0 ih"1 ih" iT n(0); T y(e); F o(Y); ?9yF o(y)g
fh"1 ih"1 ih" iT n(0); T y(e ! (e ! t)); F o(Itta)g g
Note: F o(H ) abbreviates F o(Hiroto)
F o(R) abbreviates F o(Ringo)
Having processed the entire string, the resulting tree structure must now be
turned into a complete binary tree without outstanding requirements. This
comes down to supplying the meta-variables with proper values (getting rid of
the requirements of the form ?9xF o(x)) and nding a xed tree location for all
nodes with underspecied addresses (getting rid of the requirements ?9xT n(x)).
The meta-variables providing the decorations can be replaced either through
anaphora resolution or by merging their nodes with those projected by ringo or
Hiroto; in all cases these choices may be driven largely by externally imposed,
pragmatic considerations. We have here already assumed that the nodes pro-
jected by ringo or Hiroto do indeed merge with the subject and object nodes
projected from tabeta, though this is by no means the only option. The in-
terpretation `Hiroto said that he ate the apple' can then result from anaphora
resolution, which replaces the meta-variable X annotating the subject node as-
sociated with structure projected by itta with a copy of the Formula Hiroto.
Then a completed annotation of the node decorated by F o(Y); T y(t), is pro-
vided by merging it with the sub-tree projected by Hiroto ga ringo o tabeta
The Parsing Process { A Sketch 75
to. Finally, the node requiring type t introduced by itta is xed, and steps of
Completion and Elimination then lead to all non-terminal nodes duly receiving
an annotation. One last step is the merging of the root node, T n(0) and the
node:
h" iT n(0); F o(Itta(T abeta((; x; Ringo(x))(Hiroto))(Hiroto))); T y(t)
thus nally xing all node relations in the tree. Other choices of construal of
the meta-variables could have been made to full the requirements, giving rise
to alternative interpretations.
The advantage of using the concept of (sequences of) unxed nodes in charac-
terizing a language such as Japanese is that it allows a straightforward re
ection
of the observation that such languages, in one sense, display a
at structure and,
in another, a congurational structure as languages such as English.12
In these examples we have introduced the construction of a logical form at
a leisurely pace. As the informal account of English and Japanese sentences
has indicated, though the specic lexical characterizations dier from language
to language { with dierent load carried by general rules and the lexicon { the
causality of the parsing process is essentially the same. The initial state is a
root node, with a requirement for a type t object. Each stage of the construc-
tion process corresponds to a decorated partial tree. As the tree structure is
induced, its nodes are progressively decorated with labelled terms, fullling im-
posed requirements as the words are successively scanned. Any change to a
given partial tree is an enrichment; nothing is ever cancelled en route. Only
requirements can disappear, and these only by being fullled. The nal state is
the return to the root node with no requirements left at any node. Furthermore,
the resulting structures are identical across the dierent types of language { bi-
nary trees whose types of annotation and modes of composition are common to
all languages.
It is worth noting, even at this early stage, the general way in which this
system diers from other, more familiar systems. Modelling the parsing process
as an evolving tree is not in itself new { indeed it is denitional of a parsing
system that it must dene transitions across partial descriptions of analyses of
a string (see Marcus et al. 1983). What is novel about this system, as indicated
in chapter 1, is that the system denes possible transitions from one tree to
another without reference to any independent set of axioms that constitute
the grammar of the language. To the contrary, the transition rules themselves
incorporate the grammar. They make available a concept of well-formedness for
a given language through the requirement that the words of a string contribute
in sequence to the construction of a logical form. The system will thus provide
a putative grammar formalism, our major task in this book is to explore the
viability of this idea.
12 See Hale 1983, Kiss 1991 and Diesing 1992 for the debate on whether these languages are
congurational.
76 The Dynamics of Tree Building
of words must lead from the Axiom to the nal state using each action dened
by the words in the order given, computational rules or pragmatic actions, so
that a well-formed logical form results, with no requirements on any tree node
left outstanding and no words left unprocessed.
(Axiom; w) = (T D0; w) )LCP (T D1; w0 ) )LCP : : : )LCP (T Dn; e) 2 Goal
where the constant e, as usual, denotes the empty string. As is common, we let
)LCP be the re
exive and transitive closure of )LCP .
The set of grammatical strings, given C , L and P , can then be dened as the
strings that, starting from the Axiom, will, at the last word, reach a decorated
partial tree that is either already an element of Goal or that can be moved into
Goal by applying some computational and/or pragmatic rules:
L(L; C; P ) = fw 2 L j 9T D 2 Goal : (Axiom; w) )LCP (T D; e)g.
3.2.1 Computational Rules
By denition, whenever we have (T D; w) )C (T D0; w), then we have 8w 2
L : (T D; w) )C (T D ; w), because the string is not involved in computational
0
rules. So the notion T D )C T D0 makes sense. For visual convenience, the
computational rules 2 C will be displayed as transition rules of the form
1
2
where 1 is a (schematic) tree description to which the rule is applied, and 2
is the (schematic) tree description resulting from that application: whenever
T D )fg T D0, then 1 matches tree description T D, and 2 matches T D0.16
The transitions can be grouped into a number of discrete rewrite rules. Fig-
ure 3.1 gives an overview of the rules we use in this book. The basic rules come
in pairs, one element of the pair being in a sense the converse of the other one:
one element of the pair will use the requirements as a basis for computation,
whereas the other element will generally use annotations to compute other facts.
The rst pair which we consider will make this clear.
3.2.1.1 Introduction and Elimination
The Introduction rule licenses the expansion of a requirement for an annotation
of a given type into (a subdivision) of requirements for objects of other types: it
considers a requirement for a given type and generates sub-tasks to create two
other types the fullment of which would allow us to deduce the original type
by Modus Ponens. The fact that there is a pointer involved turns this rule into
a schema whereby every choice of pointer position determines a distinct rule.
16 In any application of these rules we allow variation as to whether their eect is achieved
through lexical or computational actions.
The Parsing Process Dened 81
Introduction
f: : : f: : :?T y(Y ) : : : ; }g; : : : g
f: : : f: : :?T y(Y ); ?h#0 iZ0 ; ?h#1 iZ1 ; : : : ; }g; : : : g
where either Z0 = (T y(X ) ^ }) and Z1 = T y(X ! Y );
or Z0 = T y(X ) and Z1 = (T y(X ! Y ) ^ }):
Elimination f : : : fh#0 i(F o(); T y(X )); h#1 i(F o( ); T y(X ! Y )) : : : ; }gg
f: : : fF o(APL(; )); T y(Y ); h#0 i(F o(); T y(X )),
h#1 i(F o( ); T y(X ! Y )) : : : ; }gg
Prediction y
f: : : fX; : : :?h#iZ; }gg
f: : : fX : : :?h#iZ 0 g [ fY g; fh# 1 iX; : : :?Z g [ fY 0 g:::g
where X 2 ADR and Z = ( ^ }); Y = ;, Z 0 = ; Y 0 = f}g;
or Z = Z 0 = ; Y = f}g, Y 0 = ;
Completion
f: : : fX; : : : g; fX 0 ; : : : ; }g : : : g
f: : : fX; : : : h#i; }g; fX 0 : : : gg
where X 2 ADR; X = h#iX 0 or X 0 = h# 1 iX
and if ? 2 fX 0 : : : g then = ;:
Thinning f: : : f: : : : : :?; }g : : : g
f: : : f: : : : : : ; }g : : : g
Basic Merge
f: : : ND; ND0 : : : g
f: : : ND [ ND0 : : : g
} 2 ND0 ; (T D0 ; ND0 ) `PR V ADR(ND)
Gap Merge
( for type e) f: : : ND; ND0 : : : g
f: : : ND [ ND0 : : : g
} 2 ND00 ; T y(e0 ) 2 ND;
V ?T y(e) 2 ND0 ;
(T D ; ND ) `PR ADR(ND)
y For the meaning of ADR, see section 7.
Figure 3.1: The computational rules
82 The Dynamics of Tree Building
The Elimination rule, conversely, considers the satised sub-tasks with their
types and deduces the required type.17 The Elimination rule incorporates a form
of Modus Ponens dened under the Type label (T y), with function application
dened under the Formula predicate label (F o).
Introduction
f : : : f: : :?T y(Y ) : : : ; }g; : : : g
f: : : f: : :?T y(Y ); ?h#0 iZ0 ; ?h#1 iZ1 ; : : : ; }g; : : : g
where either Z0 = (T y(X ) ^ }) and Z1 = T y(X ! Y ),
or Z0 = T y(X ) and Z1 = (T y(X ! Y ) ^ }).
This rule sets the stage for an application of Modus Ponens under the T y pred-
icate which would give the required type.
An instance of this rule, that has been used in steps 2 and 3 of the basic
example, is the following:
f: : : f: : :?T y(t) : : : ; }g; : : : g
f: : : f: : :?T y(t); ?h#0 i(T y(e) ^ }); ?h#1 iT y(e ! t) : : : ; }g; : : : g
An application of this rule analyses a requirement for a type t object into two
new requirements, one for a type e and one for a type e ! t object. Moreover,
the pointer follows the type e requirement.
Elimination
f : : : fh#0 i(F o(); T y(X )); h#1 i(F o( ); T y(X ! Y )) : : : ; }gg
f: : : fF o(APL(; )); T y(Y ); h#0 i(F o(); T y(X )),
h#1 i(F o( ); T y(X ! Y )) : : : ; }gg
This rule applies Modus Ponens under the Type label and, in tandem, executes
function application under the Formula predicate.18
The rules Introduction and Elimination work together with the following
pair: Introduction sets the stage for applications of the Prediction rule, and the
Completion rule precedes application of Elimination.
Completion
f : : : fX; : : : g; fX 0; : : : ; }g : : : g
f: : : fX; : : : h#i; }g; fX 0 : : : gg
where X 2 ADR; X = h#iX 0 or X 0 = h# 1 iX
and if ? 2 fX 0 : : : g, then = ;.
It will be a matter of choice when to apply the Completion rule (followed by
Elimination ) in the course of a parse. In general, we will apply the rules as late
as possible. We will restrict applications of the Completion rule to situations
where all requirements in the source description have been satised. That is,
we demand that the node description containing the pointer } above the line in
the rule formulation should not contain question marks: information may only
be propagated from one node description to another if no requirements in the
source are outstanding. After all, it may turn out that some requirements are
unfulllable and thus come from a description which is not a viable description
of a tree node.
84 The Dynamics of Tree Building
3.2.1.3 Thinning
In the course of a parse, annotations and requirements accumulate within node
descriptions. A requirement within a node description ND is declared fullled
if ND contains both ? and . At this point the Thinning rule allows us to get
rid of the question mark.
Thinning
f: : : f: : : : : :?; }g : : : g
f: : : f: : : : : : ; }g : : : g
The application of some rules may require that a description no longer has
any requirements. Application of the Thinning rule must thus precede the
application of such a rule. Notice that this rule requires the presence of the exact
formula that is required. That is, we cannot apply the rule to, for instance,
f: : : ; ; : : : ; ?( ^ )g
directly. We need a derivation step which concludes ^ from these premisses
and adds this proposition to the description.
3.2.1.4 Adjunction and Merge
So far we have set out a system of rules which enables tree descriptions to be
successively enriched, starting from the initial node description. Central to this
system is the concept of partial specication: at any stage but the last, the
description of both tree nodes and tree relations may be partial. We have now
to add the rules which enable us to add node descriptions with an underspecied
tree relation to the tree description so far constructed. Recall that this, with
its use of the Kleene star operator, is the means of describing the dynamics
involved in parsing left-dislocation structures.
In order to allow underspecied tree descriptions, we need rules for their
introduction and elimination. The introduction will be taken care of by Ad-
junction rules. These introduce node descriptions which still have to nd a
proper location in the tree under development. The Elimination rule merges,
i.e. unies, two nodes in a given tree, in general a node with another whose
address is fully specied.
The general form of an adjunction rule is:
f: : : fT n(a); : : : ; }g : : : g
ffT n(a); : : : g : : : fh" iT n(a); : : : ; }gg
Addition of an unxed node.
We will use two forms of this rule. The rst form, which we call *Adjunction,
is used in the treatment of long-distance dependencies.
The Parsing Process Dened 85
*Adjunction
ffT n(a); : : :?T y(t); }gg
ffT n(a); : : : ; ?T y(t)g fh" iT n(a); : : : ; ?9xT n(x); ?T y(e); }gg
In this version a node requiring a type e is introduced, dominated by a node
with a requirement for an object of type t. This rule sets the stage for the
processing of a (sentence-initial) noun phrase which has an as yet unspecied
argument role in the structure to be projected. With this transition so dened,
left-dislocated expressions can be characterized as projecting their content in
the normal way, but now with respect to a node with an underspecied address.
Adjunction
f: : : fX; : : : ; ?T y(x); }g : : : g
f: : : fX; : : : ; ?T y(x)g fh" iX; : : : ; ?9xT n(x); ?T y(x); }g : : : g
X 2 ADR
The rule introduces a duplicate of the node currently under the pointer and
moves the pointer to this copy. The address of this new node is underspecied19
in that it is only determined that the copy lies somewhere `below' its original.
This rule sets the scene for a type x nested within a type x. That is, tree growth
by Introduction and Prediction gives now the auxiliary trees from tree-adjoining
grammar (TAG) formalisms (Joshi et al 1975), which share a label between the
root and some terminal node. If we expand the tree below the pointed node,
then we can eventually reunite the two versions of the same node by a merge,
but we might as well not have split the node. On the other hand, if we build
on top of the pointed node, then we have to construct some top node of the
same type to merge with the original one. For this we need some tree modier
of type x ! x as the function daughter of that top node. So this method of
construction allows for adjunct-like operations. In this type of construction, it is
essential that the node has been split up, because, in eect, we have interpolated
tree structure, and real interpolation is not an incremental process, and so is
debarred.
A = fT n(a); : : : ; ?T y(x); }g
B = fh"iT n(a); : : : ; ?9xT n(x); ?T y(x); }g
...
C = fh#0ih" iT n(a); : : : T y(x); }g
Here node A has generated the copy B, and C is a node which has B as an
argument daughter. Eventually C will be positioned between nodes A and
B , or could merge with A, for it is a LOFT theorem that (T n(a) ^ h#0 i> !
h#0 ih" iT n(a)).
A node description with an underspecied address has to nd a denite
location at some point in the course of the parse. For this purpose we introduce
19 In chapter 4, this will include both daughter and link relations.
86 The Dynamics of Tree Building
a general rule to combine information, that we call Merge. This allows node
descriptions with underspecied addresses to nd their xed location in a tree
description by merging with such nodes.
Basic Merge
f: : : ND; ND0 : : : g
f: : : ND [ ND0 :V: : g
} 2 ND ; T D0; ND0 `P R ADR(ND)
0
In this rule two node descriptions are combined in case the address { that is all
propositions in the ADR language of one node (ND0 above, see denition 7)
{ can be `derived' or `computed' from the pointed node using the principles in
P R. These principles allow the use of standard propositional and modal logic,
the tree-specic LOFT principles and various Vprinciples regulating the Label
and Formula predicates.20 If (T D; ND0) `P R ADR(ND), then the address
of ND is an underspecication of the address of ND0: for instance, ND has
address h"iT n(a) while ND0 has address h"1 ih"0iT n(a), or ND has address
h"1 ih" ih"0 iT n(a) and ND0 address h"1 ih"1 ih"1 ih"0 iT n(a). As an application
of the general rule we will merge an underspecied node annotated by a formula
John with a node with a fully specied location but annotated by a meta-variable
of type e (this situation will occur in chapter 6).
f: : : fh"iT n(0); F o(John); T y(e); ?9xT n(x)g
fh"0iT n(0); F o(U); T y(e); }g; : : : g
f: : : fh" iT n(0); h"0iT n(0); F o(U); F o(John); T y(e); ?9xT n(x); }g; : : : g
f: : : fh"iT n(0); h"0 iT n(0); F o(John); T y(e); }g; : : : g
To take these steps in sequence, the rst transition step is an application of the
Merge rule. By LOFT we can show,
fh"0iT n(0) : : : g `P R h" iT n(0);
so the address of one node description can be derived from the other description.
The second transition then takes account of the fact that
fh"0iT n(0)g `P R 9xT n(x) & fF o(John)g `P R F o(U);
so two applications of Thinning will give the bottom line of the derivation.
A more specialized form of this rule can be dened that requires for a merge
of two nodes that they supply complementary type information. That is, one
node should supply the type required by the other. This is a typical situation in
the characterization of left-dislocation sequences.21 We call this rule Gap Merge
because it allows a node (giving type information) to ll a gap (a node requiring
20 See section 9.2.4.
21 We retain here the familiar terminology of gap, but there are no such nodes in this system.
A gap is nothing more than a xed node with a requirement for a certain type value which in
successful derivations merges with a node with a hitherto unxed tree-node address.
The Parsing Process Dened 87
a certain type):
Gap Merge (for type e)
f: : : ND; ND0 : : : g
f: : : ND [ ND0 : : : g
} 2 ND ; T y(e) 2 ND; ?T y(e) 2 ND0 ; (T D0; ND0 ) `P R V ADR(ND)
0
Gap Merge is applied when the pointer is at a node requiring an object of type
e and the underspecied node can supply such an object. Thus this form of the
rule allows only merging of information with a gap, merging in order to full
requirements. As an example, we show how an underspecied node annotated
by a formula John may merge with a node with a fully specied location but
requiring a formula of type e.
f: : : fh" iT n(0); F o(John); T y(e); ?9xT n(x)g fh"0 iT n(0); ?T y(e)}g; : : : g
f: : : fh" iT n(0); h"0iT n(0); F o(John); T y(e); ?T y(e); ?9xT n(x); }g; : : : g
f: : : fh"iT n(0); h"0iT n(0); F o(John); T y(e); }g; : : : g
The rst transition step is an application of the Gap Merge rule as before. The
second transition is then an application of the Thinning rule with respect to the
propositions ?9xT n(x) and ?T y(e).
So Merge is the rule which has the eect of xing what has hitherto in
the parsing process been described as unxed within a tree. It is a rule that
unies two descriptions for a given node. Primary amongst the cases to which it
applies is the case when a xed node is introduced with a requirement which the
annotations of the unxed node would satisfy and there is no other candidate
for its completion. In such a case the two node descriptions which the node
may have,collapse into a description of a single node. The result will be that
the node has then a xed tree position, while nevertheless remaining consistent
with the earlier description. This rule applies in the construal of wh questions,
in topicalization structures, and in relative clauses:
(3.6) Who do you think will succeed?
(3.7) My mother, I think will succeed.
(3.8) The man that I think will succeed has already applied.
However, the rule may apply whenever there is in the current tree description
a description of both a xed node and an unxed node which could potentially
be identied; but, in most cases, this unication process will immediately lead
to inconsistency.22
The possibility of unifying a node described as unxed with one that is xed
will also allow us to unify meta-variables with their values, for a meta-variable is
22 It should be noted that a node may only be annotated with more than one Formula value
if the values stand in an update relation Fo. Recall that the Formula values re
ect the
content of the natural language expressions, so gender and number information do not project
Formula values, but values for some discrete Labels.
88 The Dynamics of Tree Building
consistent with any xed term of the same type, so nothing prevents an identical
interpretation of (3.9) and (3.10):
(3.9) My poor mother, I can just see she's going to suer.
(3.10) My poor mother, I can just see is going to suer.
Indeed, the interpretation of the pronoun in (3.9) must be construed as identical
to the annotation of the unxed node in order to lead to a well-formed outcome.
As we shall see in chapters 5 and 6, this will have considerable ramications, for
it will provide the basis for an analysis of resumptive pronouns, the distinction
between topic and focus structures, and a solution to the puzzles posed by
crossover phenomena.
3.2.1.5 Pointer Movements
The dynamics of tree constructions require actions to be taken at specic nodes
in a partial tree. These nodes are highlighted in the tree by the presence of a
pointer. Structurally this pointer is without meaning (that is, the pointer has
no role to play in the logical form under construction), but dynamically the
pointer and its location in the tree form an essential feature of a partial tree
structure. The same partial tree with the pointer at two distinct locations may
behave completely dierently with respect to actions projected by the words
of a string, or actions induced by general computational principles. In other
words, for a lexical or computational action to be eective, it must come at
the right moment.
Pointer movement can be induced in a variety of ways:
First and foremost, pointer movement is induced by the computational
rules like Prediction and Completion (see section 3.1.1, step 3).
A second source of pointer movement is given by the lexical entries. A
word projects a lexical entry, a lexical action (section 3.2.2). This action is
executed depending on a condition which is evaluated with respect to the
current pointer position, and it leaves the pointer at some new location in
the extended structure.
A last source of pointer movement is part of a general strategy on what
to do when some construction is completed; that is, if all requirements
at some node have been fullled. For instance, the pointer may then
move to some c-commanded terminal node which still has requirements
outstanding, if there is such a node.
Pointer movements are guided through the tree along tree relations. That is,
we exclude `absolute' addressing. In this way these movements can readily be
interpreted on the familiar partial tree structures. The set of movements is
constituted by the following actions. There are the basic movements, taking
one step from tree node to tree node along one of the tree relations. These basic
movements can now take part in complex ones through the regular operations
The Parsing Process Dened 89
of sequential composition `;', indeterministic choice `+', and nite iteration ` '.
Movements can also be made conditional on decorations of the starting point.
3.2.1.6 Alternative Parse Courses
The progress from Axiom to Goal is non-deterministic: at every state of the
parse, the word currently under consideration can generally be assigned more
than one structural role in the tree description by principles in C [ P . For
instance, an NP heading a natural language string may end up as subject John
admires Mary; it may be a fronted object, Mary, John admires; or it may be
a topic constituent, (As for) Mary, John admires her. The set C must contain
actions to set up these three structures. The transitions determined by the
computational rules, i.e.
(T D; w) )C (T D0; w);
are non-deterministic ()C is a multi-valued function). But, as is evident, for
the parsing process to get o the ground at all, for any given tree description T D
the set C should contain only a nite number of (alternative) actions which have
T D in their domain: only a nite number of developments can be computed.
That is, we stipulate:
ACT (T D) = f 2 C j T D 2 dom()g
is a nite set. Moreover, we should not be able to apply rules indenitely: after
a certain number of applications, we must have reached a tree description that
is no0 longer in the domain of any rule in C . Thus, for a given T D, not only is
fT D j 9w : (T D; w) )C (T D0 ; w)g nite, but also fT D0 j 9w : (T D; w) )C
(T D0; w)g.
A further action that we will describe deals with the introduction of a fresh
logical name, either a constant or some variable, to be used in some formula .
The notion of `freshness' is here dened with respect to the tree structure under
development.
freshput(a; )=
IF f"# F o(a1 )g
THEN IF f" # F o(a2 )g
THEN : : :
: : : IF f"# F o(an )g
THEN ABORT
ELSE put(F o([an =a]))
:::
ELSE put(F o([a2 =a]))
ELSE put(F o([a1 =a]))
The action freshput(a; ) generates a xed nite number of instances
of the term a and uses the rst one that is fresh in . So, to actually
contribute a fresh variable to some tree, this number has to exceed the
number of a occurrences in that tree.
This macro is used in the actions projected by proper names and pronouns.
We will now exemplify some lexical actions in detail. The name John will
have the lexical entry25
John IF f?T y(e)g
THEN put(T y(e); [#]?); freshput(John; F o(John))
ELSE ABORT
The action projected by `John' tests whether the pointer is located at a node
requiring an object of type e. If this is the case, then it annotates the node
with F o(Johni ), the rst John-copy not occurring in the tree, with the type
declaration T y(e), and with [#]?, which declares this node to be a terminal one.
If the test fails, then (this strain of) the parsing process aborts. So distinct
occurrence of the name John in one sentence will give rise to distinct logical
constants annotating tree nodes. Dierent tree nodes may end up with instances
of the same logical name, but only through a process of copying.
The English transitive verb like will have the entry:
like IF f?T y(e ! t)g
THEN make(h#1i); go(h#1i);
put(F o(Like); T y (e ! (e ! t)); [#]?);
go(h"1 i); put(?h#0 i(T y (e)))
ELSE ABORT
25 As is standard in the literature, we side-step problems in dening linguistic properties
of proper names and assume that they are logical constants. In a fuller characterization, a
linguistic name should be analysed as projecting a meta-variable in the manner of pronouns,
with restrictions on the substituend that it be a term denoting an individual suitably named.
See section 7.3.1.1 for a preliminary formulation of the analogous presuppositional properties
of denite noun phrases.
The Parsing Process Dened 93
Notice how like, at the pointed node decorated with a requirement for T y(e ! t),
releases a sequence of actions: creation of a rst daughter node, annotation of
that node with a type and a formula and the imposition on the mother node,
the predicate node, of a requirement for a second daughter of type e. Actions
decorating a node with annotations are the analogue of lexical substitution
processes in other frameworks (though it should be remembered that it is not
lexical items which are inserted in the tree in this framework, but their logi-
cal correlate). Actions decorating a node with requirements are analogous to
subcategorization statements. However, in this framework, recall, all nodes are
introduced with requirements, so unlike standard subcategorization statements
which are imposed on sisters to terminal nodes as a requirement on lexical in-
sertion in a tree, in this framework the use of requirements is more widespread.
The characterization of like illustrates how actions induced by a lexical item
may include decorations of a node other than the one at which the pointer sits.
The tense specication on a verb is another example involving annotation of a
mother node; and arguably agreement can also be expressed as an addition to
the mother node of a modal requirement. We presume this takes place from the
node requiring T y(e ! t),26 following phonological evidence27that the suÆx is
identied as entirely separate from its host (see Kaye 1996):
suÆx -s (tense marker) IF f?T y(e ! t)g
THEN put(Agr(sing)); go(h"1i);
put(T en(Si Se ); ?h#0 iAgr(sing))
ELSE ABORT
Lexical specications may also indicate the relative position in the tree of the
node currently being decorated. Thus case information, which indicates what
kind of node in the tree the formula currently being projected should be taken
as decorating, can be expressed as a modal requirement on the node under de-
scription. In Japanese, for example, in which case particles are suÆxed to an
NP constituent, the particle -o projects a requirement on the relative position
of the node currently being annotated that its mother node must be a formula
of type e ! t:28
o IF fF o(); T y(e)g
THEN put(?h"0 iT y(e ! t))); gofirst"(T n(0))
ELSE ABORT
The projection of wh expressions illustrates the need for nested specications
for a lexical item. Such nested characterizations arise in all cases where a lexical
26 We have made several assumptions about tense. First, we have assumed here that tense
projects a relation between time points. The time point relative to which the proposition
expressed is evaluated is itself a label to the propositional form in the resulting formula.
Secondly, we have assumed that tense is a decoration on a node, rather than projecting a
node of its own (see section 2.1.1.1).
27 In this book we have nothing to say about the phonology/syntax interface. So we leave
on one side the clitic-like status of -s in likes and the morphological characterization of phono-
logically reduced forms such as clitic and resumptive pronouns.
28 The suÆxes also add the information that the pointer is returned to the root.
94 The Dynamics of Tree Building
expression has distinct but overlapping conditions of use. For example, we take
wh-initial expressions in questions to project a categorial feature Q, wh-in-situ
questions to simply project the meta-variable:29
what IF f? T y ( e ) g
THEN IF fh" i?T y(t)g
THEN put(F o(W H ); T y(e); [#]?);
go(h" i); put(Cat(Q))
ELSE put(F o(W H ); T y (e); [#]?)
ELSE ABORT
Such item-internal disjunctions should not be confused with discrete lexical
entries such as might be proposed for bank1 and bank2. The item-internal dis-
junctions are for cases in which an individual word projects a number of discrete
actions, which share a core subset of properties. For example, a word in the
language may project a single Formula value with an accompanying sequence of
actions which are particular to individual structural contexts. In the example de-
ned here, wh expressions invariably project a distinguished WH meta-variable;
but in addition, when initial in a clausal sequence and annotating an unxed
node, an additional clause-typing feature Q is added to the node from which
the unxed node was introduced. Notice that this meta-variable WH is not ac-
companied by a substitution requirement ?9xF o(x). That is, the WH variables
may remain as variables in the nal structure, and wh questions are accordingly
taken to express an incomplete propositional formula, lacking a substituend for
(at least) one argument. This distinguishes them from all other lexically pro-
jected meta-variables which have to be substituted as part of the construction
process (see section 3.2.3).
The examples listed so far have primarily been English, but these should
not be taken as illustration of an invariant cross-language pattern. The details
of the actions may dier from language to language, with variation in both the
condition and the actions. We give here the projection of Arabic abil (= `like'),
showing how in subject-pro-drop languages the actions initiated by the verb
lead to the construction and annotation of the subject position and the projec-
tion of the Japanese tabeta (= `eat'),30 where the verb induces both an unxed
node requiring type t and the construction of a sub-tree all of whose nodes are
annotated:
29 Cat is a sentence-typing predicate. In general, for convenience, we shall simply use the
short-form Q (see chapter 5).
30 As elsewhere, we are here ignoring the tense suÆx -ta, taking it, for simplicity, to be part
of the stem.
The Parsing Process Dened 95
abil [Arabic] IF f?T y(t)g
THEN make(h#0i); go(h#0i); put(F o(U); T y(e); ?9x(T n(x)));
go(h"0 i); make(h#1 i); go(h#1 i); put(?T y (e ! t));
make(h#1 i); go(h#1 i); put(F o(Abil); T y (e ! (e ! t)));
go(h"1 i); put(?h#0 i)
ELSE ABORT
tabeta [Japanese] IF f?T y(t)g
THEN make(h# i); go(h# i); put(?T y(t); ?9xT n(x));
make(h#0 i); put(F o(U ); T y (e); ?9xF o(x)); go(h"i);
make(h#1 i); go(h#1 i); put(?T y (e ! t));
make(h#0 i); go(h#0 i); put(F o(V ); T y (e); ?9xF o(x));
go(h"i); make(h#1 i); go(h#1 i);
put(F o(T abeta); T y (e ! (e ! t))); gofirst" (?T y (t))
ELSE ABORT
The details of each such lexical entry in part re
ects language-particular word
order, since the specics of pointer movement determine what subsequent ac-
tions may be available. The Arabic verb abil moves the pointer back to the
predicate node awaiting a step of Prediction, hence the presence of a following
object expression, whereas the Japanese verb tabeta moves the pointer back to
the root.
3.2.3 Pragmatic Actions and Lexical Constraints
Of the types of lexical action considered so far, one important sub-type has
been omitted. There is a set of lexical actions which update the tree description
with node decorations which do not in themselves full all requirements on that
node, but serve rather as constraints on a third type of action involved in the
construction process { pragmatic actions.
There are two principal sorts of constraints on pragmatic actions:
(i) constraints on the pragmatic process of Substitution
(ii) constraints on the form of inferences derivable from the logical form once
constructed.
We have nothing to say here about the latter, apart from noting the existence
of particles such as even and only, and the connectives but, however and af-
ter all, whose distinctive property is the contribution they make not to the
truth-functional content of logical form but rather to the non-logical inferences
derivable from that form.31
Under heading (i) fall all expressions whose lexical specications systemat-
ically underdetermine their interpretation in context { i.e. anaphoric expres-
31 Recall from chapter 1 that the construction of some distinguished logical form is assumed
to be combined with some set of premisses constituting `the context', the combination yielding
some suitable array of contextually derived implications (Sperber and Wilson 1995).
96 The Dynamics of Tree Building
arg
arg pred
We can describe such a sequence with the transitive and re
exive closure of the
function daughter relation: 1 (which we can refer to in the language by h#1 i).
The arguments of the predicates can be found annotating argument daughters
along this function spine. From the perspective of an argument node (which
32 In a more complete characterization, names would also fall in this class. Unlike Pinkal
(1995) and Poesio (1996), we assume that lexical expressions with quite disjoint interpretations
are characterized as discrete lexical items. However, for present purposes, we assume that
verbs express xed concepts. See Marten (1999) for an exploration within this framework
of the view that verbs systematically underspecify their type assignment, the extractability
of NPs from within PPs being a re
ex of their optional argument status { with verbs said
to lexically project a partial representation of a predicate whose type specication, hence
content, is xed only by the structural context.
33 Constraining actions such as Substitution are where relevance and other pragmatic con-
siderations have a role to play.
The Parsing Process Dened 97
itself may dominate structure), its co-arguments can be found by moving to
the mother (upward along 0), then at least one step upward or at least one
downward along the function relation (1), and then considering the formula at
the argument daughter. This route upwards is described by the sequence CA
of modalities
CA = h"0 ih"1 ih"1 ih#0 i;
the route downwards analogously by h"0ih#1 ih#1 ih#0i.
These (sequences of) modalities allow us to formulate locality constraints on
substitution for pronouns. A substitution action SUBST (John) 2 P for the
proper name John, for instance, can now be given the (provisory) form:
SUBST (John) IF fCA(F o(John)); F o(U); ?9xF o(x)g
THEN ABORT
ELSE put(F o(John))
This substitution action checks, when the pointer is at a node annotated by
a place-holder projected by a pronoun, whether F o(John) annotates a co-
argument node. If it does, then no further action is taken; if it does not, then
F o(John) is added to the annotations of the current node.
When we turn from fully annotated binary trees to partial ones, the notion
of locality becomes somewhat more complicated. A verb projects a predicate
which is surrounded by its arguments either as already constructed or as still
required. Now the notion of a co-argument of a node includes not only the
co-arguments in the current partial tree, but also the nodes in that tree that
can become co-arguments { that is, structurally underspecied nodes.
T y (t)
John arg
U pred
In this tree, the node annotated by John is not reachable by CA from the one
annotated by U, but it could merge with the node annotated by `arg' and thus
become a co-argument. Merge is one of the actions in C , thus there is a C
successor of the above tree where John and U are co-arguments.
T y (t) T y (t)
3.3 Summary
In this chapter we have laid out a set of rules for building up annotated trees
with a propositional formula labelling its root node. In chapter 4 we shall see
how trees can be put together to yield sets of linked tree structures.
In the mean time, we should not lose sight of the underlying motivation in
developing the model. We are modelling the process of building up an inter-
pretation of a natural language string, developing a structural representation of
content by building up a tree over which denotational content can be dened.
The model re
ects the partiality of information at every stage, explicitly char-
acterizing the gap between input specications of individual lexical items and
the output content as assigned to the string. The representation will not be
established until the tree description is completed. Even the relations between
individual nodes may be relatively unxed at intermediate stages of interpreta-
tion.
The concept of interpretation that this model provides is that of an emergent
logical form induced by the progressive mapping of a natural language string
Summary 99
onto a tree structure. We don't model natural languages as inference systems,
but as devices for creating expressions over which inference can be dened.
Conceived of as the development of a grammar formalism, the current model
is also a departure from assumptions of other formalisms. The particular gram-
mar formalism set out here is not constructed according to the formal language
model. Rather, we are building in an incremental projection of structure re
ect-
ing the left-to-right sequence of words as the mode of interpretation. This makes
it very dierent from categorial formalisms along the lines of Morrill (1994) for
example. We are not dening structure for sentences by a set of rules that
merely express structural properties of strings and their interpretation; we are
modelling the building of structure as a step-by-step process that re
ects the
procedure of using the language.34 This is distinct even from those categorial
frameworks which incorporate some sensitivity to linearity into the formalism
for example Steedman 1996, Milward 1994; for in these the assumption that
natural languages are formal systems of inference is preserved as the central
methodological assumption; and the process of interpretation is, accordingly,
an incremental process of projecting model-theoretic content. In our system,
to the contrary, the incremental process is one of building the structure over
which a particular model-theoretic content is to be dened; in consequence, the
invocation of structure is ineliminable.
The present model is unlike most orthodox formalisms, in making left-to-
right projection of structure central to the explanation. The importance of this
will emerge in the course of the next chapters. Unlike most other formalisms,
moreover, specically HPSG (see Pollard and Sag 1994) and LFG (see Bresnan
1982), there is no structure dened over the string itself. A logical form paired
with a string is, rather, a progressively projected representation of a content for
that string.
In dening the interpretation process in irreducibly structural terms, the
model is similar to Discourse Representation Theory (DRT; see Kamp and
Reyle 1993). Conceptually, it is very close to DRT, sharing with that paradigm
a commitment to modelling the process of building up interpretations. How-
ever, unlike DRT, the sequencing intrinsic to the process is not merely sentence
by sentence, but word by word. The interpretations, once completed, have a
bottom-up compositional semantics, as does a DRS (discourse representation
structure). The dierence between the two formalisms is that in our formalism
it is the left-to-right process of building up the structure which itself provides the
basis for syntactic explanation. And the partially top-down process of structure-
building, far from being problematic for the preservation of compositionality,35
is fully consonant with the tree growth concept of compositionality advocated
here.
In dening the central role of tree growth, there are some similarities with
the Minimalist programme (in particular, given the Minimalist Programme rule
of Merge that induces growth of tree structure, see Chomsky 1995); but in the
34 For a discussion of this distinction in comparing logic formalisms, see Gabbay and Hod-
kinson 1992.
35 For discussion of compositionality in DRT see van Eijck and Kamp 1997.
100 The Dynamics of Tree Building
103
104 Linked Tree Structures
prime example is the interaction between anaphora resolution and the process
of establishing a xed position in a tree for an initially unxed node in the
interpretation of relative clause sequences. The substantive interest of such an
extension is that it involves cross-language variation. Despite the universality
of concepts of long-distance dependency and anaphora resolution, we get con-
siderable variation across languages as to how these two processes interact. As
we shall see shortly, it is by paying close attention to what information is made
available at each step of the tree development that we shall have a systematic
basis for characterizing the range of cross-linguistic variation. In this chapter we
set out an account of how relative clauses are parsed, showing how the account
to be developed for one language can be extended, with minor variations, to
characterize the multiplicity of forms displayed across languages.
Our next formal move, then, is to extend the tree description language so as
to allow multiple trees associated with a single sentential sequence { linked tree
structures. The concept of linking tree structures involves the projection of two
tree structures in tandem and a
ow of information from one tree to another.
To take the most transparent case by way of example, consider informally the
step-by-step process involved in processing:
(4.1) John, who I like, criticized Hegel.
The building up of an interpretation for this sequence involves building a pair of
structures { one for John criticized Hegel, a second for who I like. This process
is engineered by the relativizing element who, which uses the representation of
John in the rst structure to construct a copy of this representation in the sec-
ond structure, thus securing a `link' between the two structures in the form of a
shared formula. This quasi-anaphoric process associated with an initial wh ele-
ment leads to the presence of an unxed node in the second tree whose position
then has to be resolved { exactly as in left-dislocation sequences considered in
chapter 2. This sequence of actions apparently required for English is, however,
not the only way of securing appropriately linked tree structures, and it is the
dynamics of how such linked tree structures are incrementally projected from
which the dierent types of relative structure and their associated restrictions
will emerge. As we shall see, the initial underspecication of the tree relation
between the node projected from a wh expression and those of its contiguous
expressions in the string remains central to the account. It is just that there is
something else going on in addition { the process of incrementally building up
linked tree structures.
the head (as in English), in some with the modifying structure preceding the
head (Japanese, Hindi, Chinese), in yet others with part of the head preceding
the modifying structure, part of it following (Chinese):
(4.2) Hanako ga kaita hon [Japanese]
HanakoNOM writeP AST book
`the book that Hanako wrote'
(4.3) ai ta muqin de nei ge ren [Chinese]
loves his mother rel that classifier man
`that man that loves his mother'
(4.4) nei ge ai ta muqin de ren
that classifier loves his mother REL man
`that man who loves his mother'
In some languages (German and optionally English), the relative structure is ex-
plicitly introduced by an anaphoric-like expression, either wh (English), demon-
strative (Dutch and German), or denite (Arabic) in form, carrying over infor-
mation from the head node into the modifying structure:
(4.5) l- mudarris illi mabsu:t [Egyptian Arabic]
the teacher that happy
`the man that is happy'
As we shall see in due course, these languages vary as to how strong this
anaphoric property is, this correlating with obligatory presence or absence of a
resumptive pronoun. Thus in Egyptian (and other forms of) Arabic, all non-
subject positions in a relative require a resumptive pronoun:
(4.6) l- mudarris illi Magdi darab-u
the teacher that Magdi hit him
`the teacher that Magdi hit'
In other languages (Hebrew) the complementizer may not carry any structure-
specic encoding { it is the general subordinating device that is used to mark
the relative (note the two occurrences of se in (4.7)):
(4.7) ha-?-is se xasavt
p se Dani pagas (?oto)
p
over the re
exive transitive closure of the daughter relation (#) alone.
For the dynamics of the tree construction, we dene a variety of rules which
all have in common that, at some point in the construction process, a second tree
is initiated by introducing a top node with requirement ?T y(t), which cannot be
completed without there being some node annotated by a given formula from
the rst tree. A rule of this type is called a LINK Introduction rule. As we
shall see, languages vary in the way the constraint of a shared formula in a
pair of trees is implemented; indeed, we shall also see variation within a single
language.
4.2.1 Dening Linked Tree Structures
The rst step is to formulate the extensions to the structures and the language
we have dened in chapter 2, which are required to deal with the notion of linked
trees.
Denition 11 (Link Relations and Link Modalities) A Linked Basic Tree
structure LBT R T is a nite set of partial trees T1; : : : ; Tn with disjoint Tree
Domains and a Link relation L S1in T rDomi S1in T rDomi such
that, if n L n0 holds between n 2 T rDomi ; n0 2 T rDomj , then i 6= j . More-
over, there is no loop Ti L : : : L Ti in T (where Ti L Tj , if there are
n 2 T rDOMi ; n0 2 T rDomj : n L n0 ).
The language DU has modalities hLi and hL 1i with interpretation
T ; n j= hLi i 9n0 : n L n0 & T ; n0 j=
T ; n j= hL 1 i i 9n0 : n0 L n & T ; n0 j=
On the basis of the Link relation we dene the relation D as the re
exive and
transitive closure of [ L and introduce the modalities hDi and hU i with
the denition
T ; n j= hDi i 9n0 : n D n0 & T ; n0 j=
T ; n j= hU i i 9n0 : n0 D n & T ; n0 j=
Now, a Linked Partial Tree Structure is a structure that can be completed to
a Linked Basic Tree Structure. Because the Link relation L is a fully speci-
ed tree relation, we will expand the domain of the tree node predicate T n to
include values including a link step.
hL 1 iT n(a) () T n(aL),
so when a LINK structure is built from some head node with address T n(a),
the root node of that LINK structure can be identied as T n(aL).
Given a node n in a Linked Partial Tree Structure T , we call the local tree
with respect to n the set of nodes fn0 2 T j n i n0; i 2 f#0; #1; #; #gg. That
is, the local tree with respect to node n is the set of nodes that can be reached
without crossing a Link connection. On the other hand, the global tree with
respect to a node n 2 T is the set of nodes that can be reached from n by i,
where i may be any sequence of indices.
The Analysis { A Sketch for English 111
4.2.2 Relativizers Annotating Unxed Nodes
Of the analyses we give in terms of linked tree structures, the rst is English,
in which the relative element is a pronominal-like device, annotating an unxed
node constructed within the newly induced LINK structure. We take non-
restrictive relatives (4.21) as providing the clearest form of the pattern we want
to analyse. In (4.21), the formula projected from John is copied over into the
tree associated with who I much admired, leading to an interpretation of (4.21)
as equivalent to (4.22):
(4.21) John, who I admire, has left MIT.
(4.22) John has left MIT, and I much admire John.
The copying process is ensured by the relative complementizer. These com-
plementizers, appropriately called relative pronouns by Jespersen (1961), are
dened to transfer information from one structure to another. The rule of
LINK Introduction applies to a node in a tree decorated by some declarative
unit h: : : ei : and induces the root node of a new tree in whose subsequent
development a copy of is required to occur. The resulting relation between
this node and the new root node is the LINK relation, the tree so introduced
the LINKed tree.4 This new tree has, by denition, a requirement imposed on
it not merely to derive a formula of type t, but more specically to establish
within it a node at which a copy of occurs.
In English we have to ensure the initial position of the wh expression in the
clausal sequence, albeit possibly embedded within a containing constituent, so
we dene the LINK Introduction rule for this language as creating both the
root node of the LINK ed tree and an unxed node requiring a copy of the head
formula, so that ensuing development of the LINK ed tree may follow the same
sequence of transitions as displayed in left-dislocation structures. On this rst
formulation, we assume that the unxed node, when the requirement of con-
structing a copy of the head is fullled, will be fully annotated:
z
head
}| {
f: : : fX; F o(); T y(e); }gg
f: : : f| X; F o({z); T y(e)g}; fh
|
L 1iX; ?T y(t)g; fh"ihL 1 iX; ?F o(); ?T y(e); }gg
{z } | {z }
head linked node unxed node
The eect of this LINK Introduction rule applied in the analysis of (4.21)
is as follows. It takes the partial tree resulting from having parsed John and
creates the structure in gure 4.1.5 Note the imposition of the requirement
4 We shall refer to the resulting pair of tree structures as a pair of linked structures.
5 As before, throughout this chapter, the tree displays are essentially informal, and the
complete description is given only in the list below.
112 Linked Tree Structures
fT n(a); ?T y(t)g
z
head
}| {
f: : : fX; F o(); T y(e); }gg;
f: : : f| X; F o({z); T y(e)g}; fh
|
L 1iX; ?T y(t)g; fh"ihL 1 iX; ?T y(e); ? # F o(); }gg
{z } | {z }
head linked node unxed node
According to this specication, LINK Introduction induces an unxed node
with a requirement for a copy of the head formula to occur in a dominated
position. The projection of the relative pronouns which and whom which oc-9
cur in pied-piping constructions will, accordingly, have to be more complicated:
7 Notice that the value of the Formula predicate is bound to the variable x, as dened in
chapter 3, denition 10.
8 Recall the distinction between internal and external uses of `" ', in which the latter is
interpreted over the dominance relation, and the former over the relation which need not
(yet) be the re
exive transitive closure of the immediate dominance relation.
9 We ignore here the peripheral use of which as a relativizing determiner, for simplicity:
(i) ??I read the book, which monstrosity I refused to circulate.
However, in principle, a lexical characterization of which as a determiner is unproblematic.
The complication which its denition presents is analogous to the determiner the, which is
dened to take the nominal meta-variable and lift it to become a formula annotating the root
node of the local tree of type e (see section 7.3.1.1).
114 Linked Tree Structures
f?T y(e)g
fhL 1 iX;
?T y(t); }g
with the restrictor X (X; Man(X )) and one annotated by some fresh nominal
variable U. This second node is LINK ed to the top node of a second tree, and
the nominal variable is transferred as usual. With due processing of the following
relative clause, the formula F o(Like(U)(Sue)) annotates the root node of the
LINK ed structure. Now, the conditions for combining the nominal content with
that of the LINK ed structure are met, yielding a formula of type cn:
F o(U; Man(U) ^ Like(U)(Sue))
annotating the head node. This formula will then combine with the formula
projected by the determiner to yield a term of type e (see chapter 7 for more
details of our account of quantication); and the nal tree is as schematically
displayed in gure 4.6 (in which all details other than F ormula values are
suppressed).
The account of English relatives proposed allows a very natural, though
language-idiosyncratic, extension to complementizer-less relative forms. All that
is needed is the option of allowing a copy of a formula from the host node to be
transferred over not as a requirement, but as an annotation:
LINK-COPY Introduction
z
head
}| {
f: : : fX; F o(); T y(e); }gg
f: : : f| X; F o({z); T y(e)g}; fh
|
L 1iX; ?T y(t); }g; fh" ihL 1 iX; F o(); T y(e); gg
{z } | {z }
head linked node unxed node
118 Linked Tree Structures
F o(Ignore(the,x; Man(x) ^
Like(x)(Sue))(John))
o((U; Man(U))^
F o(P (the,P )) FLike (U)(Sue))
F o(U) F o(Man)
F o(Like(U)(Sue))
F o(Sue) F o(Like(U))
F o(U) F o(Like)
Figure 4.6: The nal result of processing (4.26) with restrictive relative
Notice that with the process of LINK dened as annotating the unxed node
with a copy of the formula of the host node, the eect in these structures is
exactly that of an empty complementizer.14
This analysis of relative clauses has two distinctive properties. Firstly, the
tree structure corresponding to the relative sequence is projected externally to
14 This characterization of complementizer-less relatives makes no reference to whether or
not the head node is part of a quantied phrase. It might be argued that such complementizer-
less forms are exclusively restrictive in construal. However, Swinburne (1999) argues (from
a slightly dierent formulation of the LINK transition rule) that examples such as (i){(ii)
involve the projection of a LINK ed structure for the small-clause predicate:
(i) A man came in late.
(ii) Everyone went home tired.
In particular, he points out that these have an E -type form of interpretation, requiring that
the entire term annotating the head is copied onto a LINK ed structure as the subject of the
small-clause predicate. Such examples also provide evidence against the observation of Fabb
(1990) that non-restrictive relatives do not modify quantied noun phrases.
The Analysis { A Sketch for English 119
the tree projected by the main clause.15 Secondly, the relativizing element is
pronominal in nature rather than quanticational; and the characterization of
relative pronouns as quasi-anaphoric captures both the similarities between rel-
atives and wh questions, and their dierences. Like wh expressions in questions,
relative pronouns project an annotation for a node whose nal position in the
tree is not yet xed, but unlike wh expressions in questions, which project a
meta-variable (W H ), relative pronouns serve as a channel, providing a copy of
the formula annotating the node from which the LINK ed structure is initiated.
Furthermore, unlike wh questions, there is no wh-in-situ relativizing pronouns
{ all relative pronouns must project a formula onto an unxed node. Notice,
incidentally, that there is no discrete logical type corresponding to the category
of complementizer. A further advantage of this analysis is its articulation of
the dierence between restrictive and non-restrictive relatives as simply the dif-
ference between constructing a copy within the independent structure of either
the formula annotating the root node of the sub-tree projected by some noun
phrase sequence or a nominal variable which is contained within it.16 The ar-
ticulation of this dierence does not require any invocation of structures licensed
15 There are a number of ways in which this LINK relation might be used for more general
application. First, given that the modal operator hLi is dened merely in terms of a relation
between nodes in a tree, we would expect also to dene LINK operations which involve no
transfer of information from the one tree to the other. Though we shall not pursue this
here, this provides a natural starting point for an account of co-ordination, which would have
the benet of debarring Merge applying to unify unxed nodes with some node in a second
conjunct of a noun phrase as in:
(i) *Who did John see Mary and?
Second, we might extend the present denition of the LINK Introduction process to allow
transfer of formulae of arbitrary type, imposing a requirement for a second occurrence within
that newly induced LINK ed tree of formulae of various types. Such an extension would allow
the following data to fall within the general rule:
(ii) The man was slovenly, which his brother was not.
(iii) John gave a book to Mary, to whom we hadn't expected him to give anything.
(iv) The man failed to clean up, which was annoying.
The generality of this phenomenon turns on a combination of factors { the process of LINK
itself; what types of expression are licensed in an unxed position at the projection of a
root node in a tree; the type of the expression triggering the transfer; and the allowed type
ambiguity of the relative pronoun, which in English can be one of several types { e, cn ! e,
e ! t, or even type t. Though we do not explore this in detail here, it is arguably the
intersection of these which gives rise to the apparent restriction on what can be transferred
across a LINK relation, not any restrictiveness intrinsic to the LINK process itself.
16 Sag (1997) argues that while complementizer-less relatives are N0 adjuncts, relatives with
an explicit complementizer are instances of NP adjunction (and hence requiring the analysis
we have given here to non-restrictive relatives), on the basis that restrictive relatives with
a complementizer can be constructed on nominals that have no obvious internal analysis in
specier and N0 :
(i) All who lost money were reimbursed.
However, nominal ellipsis is independently warranted for these constructions, given the well-
formedness of (ii), suggesting that at the level of interpretation the LINK relation is never-
theless dened over nominal variables:
(ii) All were reimbursed.
120 Linked Tree Structures
T y (e ! t)g
fF o(U), fF o(Mudarris);
T y (e)g T y (e ! cn)g
fF o(Darab), fF o(U),
T y (e ! (e ! t))g T y (e); }g
Figure 4.7: The relativizer as projecting a requirement, satised by a subsequent
pronominal
mudarris
IF f?T y(e)g
THEN IF f#1 9xF o(x)g
THEN go(h#0 i); make(h#1 i); go(h#1 i);
put(F o(Mudarris); T y (e ! cn); [#]?);
go(h"1 i); make(h#0 i); go(h#0 i); put(F o(V); T y (e))
ELSE make(h#1 i); go(h#1 i);
put(F o(X (; x; X )); T y (cn ! e));
go(h"1 i); make(h#0 i); go(h#0 i); put(?T y (cn));
make(h#1 i); go(h#1 i); put(F o(Mudarris)),
T y(e ! cn); [#]?);
go(h"1 i); make(h#0 i); go(h#0 i); put(F o(V),
T y(e); Def ( ))
ELSE ABORT
illi
IF fF o(x); T y(e); Def (+); [#]? g,
THEN make(hLi); go(hLi); put(?T y(t); ?hDi(F o(x); T y(e)))
ELSE ABORT
l
IF f?T y(e)g
THEN make(h#1i); go(h#1i); put(F o(P (the(x); P (x))); T y(cn ! e));
go(h"1 i); make(h#0 i); go(h#0 i); put(?T y (cn));
make(h#0 i); go(h#0 i); put(?T y (e); Def (+)); gofirst" (?T y (e))
ELSE ABORT
Note the lack of an unxed node in the LINK Introduction rule for Arabic and
the use of the hDi modality, interpreted over the re
exive transitive closure of
L [ . With ?hDiF o(U) on the root node of the linked tree, the LINK
Introduction displays the second respect in which languages such as Arabic
dier from Romance and Germanic languages: the copy of the head formula
may occur anywhere in the global tree, not just in the tree local to the top
node.
Notice, next, the characterization of mudarris (= `teacher'). Item-internal
disjunctions of this sort are used to express dierent eects which a lexical item
may project in dierent structural contexts.
Then, with the lexical specication of illi (and the determiner l ), the dif-
ference between English and Arabic resumptive use of pronouns is ensured.
Locality apart, the primary dierence between the two types of language is that
in English the relative pronouns are properly anaphoric in nature, securing a
copy of the head formula within the LINK ed structure. Illi, by contrast, does
not { it merely initiates a LINK introduction, imposing a requirement on the
LINK ed tree that it contain such a copy.
This cross-linguistic dierence in relativizers might seem surprising, particu-
Relatives: Towards a Dynamic Typology 125
larly given that illi is a depleted form of demonstrative. However, the intrinsic
weakness of illi { merely inducing the construction of a LINK ed structure in
which a copy of the formula annotating the head should occur { is not such a
surprising property if we recall that there are two dierent types of anaphoric
devices in language, the regular anaphoric form and more specialized expletive
forms. Expletive anaphoric expressions have a range of eects which mark them
out as distinct from their non-expletive counterpart. Their interpretation is pro-
vided not from what is already made available within the discourse or23structural
context, but from some expression which occurs only subsequently:
(4.34) Iti is likely [that John will come]i.
(4.35) Therei arose [a storm]i.
Expletive uses of pronouns are not an option freely available for all pronomi-
nal forms, but have to be lexically dened as a separate use of each appropri-
ate pronoun. They are, furthermore, characteristically associated with specic
structural contexts. For example, indenite expletives in English occur only in
subject position and only with a restricted subset of predicates; and it expletives
are obligatorily associated with sentential forms and occur only with a restricted
set of verbs. In our analysis, one essential feature shared by all expletives is that
they do not supply annotations; rather, they project requirements: and these
have the eect of ensuring the presence of some later sequence of words which
can be construed as satisfying them. For instance, in Romanian, clitic pronouns
function like agreement devices, projecting requirements with which some sub-
sequent expression must agree (the following examples are from Dobrovie-Sorin
1990):24
(4.36) len nu (i)-am examinat pe multi elevi
yesterday not (them)-past1:Sing examined pe many students
`Yesterday I did not examine many students.' [Hungarian]
(4.37) invita-(l) pe oricare elev
invite-(himACC ) pe any student
`Invite any of your students.'
(4.38) no (i)-am scris nimanui
not (himDAT )-past written nobodyDAT
`I haven't written to anybody.'
Looked at from this perspective, illi becomes less puzzling. It is not so much a
relative subordinator which happens to project agreement properties that have
to be simply stipulated. Rather, it is an expletive relative pronoun that initiates
23 The co-indexing indicates correlation of interpretation between two types of expression,
not any properties of denotation.
24 Note that there is no presumption of deniteness in these Romanian examples, so these
clitic pronouns cannot be taken to be indexical and indicative of some independently identi-
able referent.
126 Linked Tree Structures
the introduction of a LINK transition, and leads to the decoration of the root
node of the LINK ed tree with the requisite formula requirement. This speci-
cation, in virtue of being a requirement, forces the later occurrence of some
expression to provide the appropriate value. In this respect it is like an agree-
ment marker, except that the requirement,25 though lexically projected, does not
have to be satised within the local tree. This property is common amongst
relative complementizers (see also the complementizer marking of Irish { Mc-
Closkey 1988, DuÆeld 1995), a phenomenon forming part of the general pattern
of relative complementizers as encoded anaphoric devices. Indeed, it is one of
the advantages of using a modal tree-description language that such statements
of locally projected requirements to be satised non-locally are straightforwardly
available.
From this account, one might expect that, just as individual languages dis-
play both non-expletive and expletive forms of the same pronominal, subject
to lexical variation, so an individual language might contain both an expletive
and a non-expletive variant of a relativizing particle. Thus Hebrew has both
a relative clause strategy with `gaps' and one with free use of resumptive pro-
nouns, and can be dened as having both the anaphoric-like form of relativizer
analogous to the empty relativizer in English and the expletive26 form of rela-
tivizer analogous to the relativizing complementizer in Arabic. It is notable
that the gap variant is subject to subjacency eects, whereas the resumptive
form is not:
(4.39) ha-?-is se xasavt
p se Dani pagas (?oto) p [Hebrew]
the man that thought2:F em that Dani met (him)
`the man that you thought Dani met'
The prex se introducing the relative has no morphological re
ex of anaphoric-
ity. To the contrary, it is simply a general subordinating device. Notice the
identical form of the prex se for both relativizing and subordinating particle
and the quite dierent form of the pronoun ?oto. p
the man that I think that about him you said that Sara
katva sir
wrote poem
(4.41) ha-?-is
p se ?alav
p ?ani xosev se
p ?amarta se
p Sara
the man that about him I think that you said that Sara
katva sir
wrote poem
(4.42) ha-?-is
p se ?ani xosev se
p ?amarta se
p ?alav
p Sara
the man that I think that you said that about him Sara
katva sir
wrote poem
`the man that I think you said that Sara wrote a poem about him'
As these examples show, a pronominal prepositional phrase can be `fronted' in
Hebrew to any position associated with a clausal boundary occurring between
the head and its point of construal in the emergent structure. This is a process
identied in the literature as a separate process called left-dislocation (see section
4.3.2 and Anagnostopoulou 1997).
Notice how such an update proceeds in steps. In the rst step, an unxed
node is introduced, and the initially imposed requirement is satised (through
*Adjunction and anaphora resolution).27 In the second step, the tree node
identication for that node gets xed (through Merge ). We have the progression
from
fX; ?T y(t); ?hDi(F o(); ?T y(e))g
to
ffX; ?T y(t)g : : : fh" iX; F o(); T y(e); ?9xT n(x)gg
as the rst step, with the construction and annotation of an unxed node, so
that only the identication of the tree node remains as a requirement to be
fullled; and the progression at the second step from
ffX; ?T y(t)g : : : fh" iX; F o(); T y(e); ?9xT n(x)gg
to
ffX; ?T y(t)g : : : fY; F o(); T y(e)gg
27 We assume here for simplicity that the form ?alav is of T y(e), the prepositional content
p
where the tree node is nally xed at Y . No special process needs to be dened
to characterize this sequence of steps { it is merely a combination of processes
that are available independently.
4.3.2 Variation in Locality
As we have seen, the Semitic family of languages displays variation as to whether
the relativizing element is or is not expletive-like. Independently, this element
also displays a lack of any locality restriction on the second occurrence of the
formula annotating the head. Since these two parameters are independent, we
would expect there to be languages in which a relative pronoun projects merely
a formula requirement on the root node of the LINK ed structure (as in Arabic)
but under the restriction that, as in English, the copy of the head formula
occur within the local tree. And indeed, between Romance languages we get
variation relative to a shared locality restriction { for example, Italian and the
closely related language of Romanian (see Dobrovie-Sorin 1990).28 In standard
Italian, resumptive pronouns do not occur in ordinary relative constructions
(see Cinque 1990). A gap is required: the various relative complementizers, as
in English, ensure the presence of the formula from the head in the LINK ed
tree.29 However, in Romanian, whose clitic pronouns have both expletive and
non-expletive functions,30 the relative particle, care, is obligatorily paired with a
subsequent resumptive pronominal, much as in Arabic. However, unlike Arabic,
this Romanian pronominal is subject to strong island restrictions, and may not
occur within a relative clause or a clausal adverbial (the data are from Dobrovie-
Sorin 1990):
(4.43) baiatul pe care l- am vazut [Romanian]
the boy pe which him have1:SING seen
`the boy that I saw'
(4.44) *baiatul pe care am vazut
the boy pe which have1:SING seen
`the boy that I saw'
28 There is widespread variation between the dierent dialects of Italian in this respect across
the dierent uses of relative and non-relative pronouns: see Escobar 1997.
29 There are two types of complementizer: the invariant che, the general subordinator, that
does not allow pied-piping (there is also a form cui which occurs with prepositions), and Det
+ quale which is a pronominal form occurring only with pied-piping forms (see Cinque 1995,
Escobar 1997). These two patterns correspond to the two LINK rules postulated, LINK
Introduction and LINK-COPY Introduction. Cinque (1995) reports that in `very colloquial'
Italian, the invariant che is used with free availability of either gap or resumptive pronoun:
see chapter 5 for discussion of how anaphora resolution and Merge may interact in such cases.
30 In our characterization of expletive, the clitic pronouns in (i) are putative expletives:
(i) L-am vazut pe Ion.
him-(we) have seen pe John
`We saw John.'
Relatives: Towards a Dynamic Typology 129
(4.45) *Omul pe care cunosc femeia care l-a
the manj pe whichj (I) know the womani whichi himj have
^int^ilnit
met ej
*'the man which I know the woman who met.'
If we dene care as an expletive relativizer projecting a requirement ? # F o()
on the root node of the LINK ed structure for an occurrence of the requisite
formula within the local tree, rather than the weaker form ?hDiF o(), we re-
ect the dierence in distribution of resumptive pronouns in Italian and Roma-
nian: only the latter requires (strong-island-respecting) resumptive pronouns
in relatives. In Spanish we have a mixed system much like Hebrew, in which
resumptive pronouns are in some positions optional with the neutral que form,
and in others (the oblique positions) obligatory.31
In sum, dierent forms of relative clauses vary according as a requirement of
a copy of the head formula in the LINK ed structure is or is not realized immedi-
ately as an annotation on some introduced unxed node. In the absence of any
such realization, the outstanding requirement will necessitate the occurrence of
a pronoun, for only this can secure the presence of the required copy. But in
all variants, the imposition of the restriction that leads to two copies of a single
31 Suner's (1998) analysis of resumptive pronouns in relative clauses is supercially close to
our own, with the distribution of resumptive pronouns due to whether the complementizer
does or does not have strong pronominal features. However, Suner explicitly sets aside re-
sumptive pronouns as being a mere rescue device, and does not include them within a general
characterization of anaphora. Her Minimalist analysis takes left-dislocation structures to be
an imperfection, motivated only by some morphological feature needing to be checked. This
pronominal feature is not an interpretable feature, and is deleted upon checking, playing no
part in construal at the LF interface. Specically, it is not this feature which ensures the
co-indexing with the head, but the predicational `force' of the complementizer. Relative to
these assumptions, resumptive pronominals are the preferred form of analysis, providing a
derivation in which nothing moves. The rst problem for this analysis is that she has to allow
a grammar-internal process of `lexical manipulation' which ensures that the relative pronoun
if not moved surfaces as a regular lexical pronoun. It is by no means obvious how this pro-
cess is to be dened: as she herself observes, the resumptive element may also be realized by
some anaphoric expression other than a pronoun, data which are problematic, given that the
replacement of the relative pronoun at PF by the pronominal resumptive is licensed in virtue
of adding nothing to the interpretation:
(i) The boys who I could see both were shaking with fear, ran away from the bomb.
(ii) There was discussion about infrastructure which I could see he wasn't going to take that
aspect into account.
Third, to re
ect the observation that relative wh elements contrast with wh interrogative ele-
ments, Suner proposes that anaphoric operators are restricted so as not to appear as variables
at PF, but are subject to a grammar-internal process of manipulation to ensure their pronom-
inal form. But this stipulated restriction is nothing more than a statement of the problem.
It makes it all the more puzzling that the expressions replacing such relative pronoun forms
are all anaphoric. In the light of this, the assertion that anaphoric operators cannot remain
in situ, because `a relative pronoun cannot behave as a bound variable at PF0 is mysterious:
it is referential expressions such as these which are taken not to require movement. More
generally, this account does not explain why wh forms should be quasi-anaphoric expressions,
why they should allow variation, or why they should interact with anaphora resolution to
yield the assigned interpretations.
130 Linked Tree Structures
term in the two structures can be locally dened as part of a LINK introduction
rule.
4.3.3 Topic Structures and Relatives
This analysis of relatives extends naturally to so-called clitic left-dislocation
structures (see Anagnostopoulou et al. 1997), enabling us to bring out the
parallelism displayed by these two types of structure in head-initial languages
in a straightforward way. In clitic left-dislocation structures, there is a pairing
between a left-dislocated constituent and a pronominal in the main structure,
with the dislocated expression characteristically separated from the remainder
of the sentence by a marked intonational break. Re
ecting this, we analyse clitic
left-dislocation structures as a pair of linked structures: the initial noun phrase
sequence is taken to project a tree with a root node annotated by some formula
of type e, and that node has the root node of a second tree LINK ed to it,
with the familiar constraint imposed on the development of this new structure
that it contain some node annotated by F o() (see gure 4.8):
(4.46) Magdi, Ali abil-ha [Egyptian Arabic]
Magdi, Ali met-her
`Magdi, Ali met.'
(4.47) irragil, huwwa mabsu:t
the man, he happy
`The man, he is happy.'
(4.48) Gianni, lo vedro domani [Italian]
`Gianni, him I'll see tomorrow.'
(4.49) Ion, l-am ^int^ilnit anul trecut [Romanian]
John, him-(I) met year last
`John, I met him last year.'
(4.50) As for John, I like him. [English]
`Shalom, I think that you said that Sara wrote a poem about him.'
(4.60) Shalom, ?alav
p ?ani xosev se
p ?amarta se
p Sara katva
Shalom, about him I think that you said that Sara wrote
sir
poem
`Shalom, I think that you said that Sara wrote a poem about him.'
(4.61) Shalom, ?ani xosev se
p ?amarta se
p ?alav
p Sara katva
Shalom, I think that you said that about him Sara wrote
sir
poem
`Shalom, I think that you said that Sara wrote a poem about him.'
On the other hand, in languages in which the relativizing element is a relative
pronoun that presents the copy of the head formula as annotation to some un-
xed node in the LINK ed structure, there should be no such parallelism, and a
resumptive pronoun should be required only in clitic left-dislocation structures.
And so it is that Romanian and Italian dier, with Romanian patterning, like
Arabic, with perfect symmetry between topic and relative structures, but Italian
displaying asymmetry in the two forms of structure. Italian, despite precluding
resumptive pronouns in relative structures, requires a resumptive pronoun in
clitic left-dislocation structures, where it is characteristically associated with
so-called comma intonation, as in Romanian (see Cinque 1990, Dobrovie-Sorin
1990):
(4.62) *l'uomo che l'aprezzavamo [Italian]
the man that we appreciated him
`the man that we appreciated'
(4.63) Gianni, l'apprezzavamo.
Gianni we appreciated him
`Gianni, we appreciated.'
(4.64) baiatul pe care l-am vazut [Romanian]
the boy pe which him-have1:SING seen
`the boy that I saw'
(4.65) *baiatul pe care am vazut
the boy pe which have1:SING seen
`the boy that I saw'
(4.66) Ion l-am ^int^ilnit anul trecut.
John him-have1:SING met year last
John, him I met last year.'
Relatives: Towards a Dynamic Typology 133
Notice, incidentally, that clitic dislocation involves the construal of the clitic
pronominal from a dislocated hence unxed position. In clitic left-dislocation
structures then, in Italian and Romanian alike, the process of interpretation
echoes that associated with the previous Hebrew left-dislocation data { with
a LINK transition followed by the projection of information from a pronoun
onto an unxed node being used to full the modal requirement imposed by the
LINK transition.32 What this display of cross-language variation shows { both
in relative clauses and in left-dislocation structures { is the immediate benet
of the analysis in terms of the building of linked tree structures. For in each
of the analyses { in English, in Arabic, in Italian, and in Romanian { the only
variation that has been manipulated is whether the relativizing element has
pronominal-like properties or not and the dierence in the locality requirement
associated with the node that is to be eventually annotated by the the copy of
the head formula.
LINK ed structure. And so on. There is also the possibility of either one be-
longing to a subordinate clause. It may not be possible to resolve these choices
until the appropriate cluster of verb+nominal or verb+verb sequence becomes
available.
We have assumed that interpretation in Japanese is only very weakly goal-
driven, with case specications narrowing down the space of possible tree posi-
tions for the node they decorate, the driving force of interpretation being indica-
tors of structure such as verbs (which occur nally in any clausal sequence) and
morphological indicators of boundaries between one structural domain and an-
other (such as indicators of relative or other clausal boundaries). Let's consider
the following example containing a restrictive relative clause.
1 f?T y(t)g
fF o(U),
T y (e); f?T y(e ! t); }g
?9xF o(x)g
fF o(V), fF o(Muita),
T y (e);
?9yF o(y)g T y(e ! (e ! t))g
1 = fF o(Hiroto); T y(e); ?9xT n(x)g
Here we will use the variable projected as the object argument of Muita in
preparation of the intended interpretation, though this is a matter of choice,
a pragmatic action like other anaphoric choices. The problem here is that the
value of the term to be shared between the two nodes is not yet known at the
point of the transition; the variables present have to be identied, through the
rule of Japanese LINK Introduction. We display this as:
f?T y(t)g f?F o(V); }g
1 fT y(t)g
So, from the pair of a top node of some tree with a Formula value of type t and
some arbitrarily selected local node still requiring a xed value (i.e. annotated
with a meta-variable), the rule licenses the introduction of a LINK ed node
whose Formula value must, by denition, be that of the selected one. As a
result, in the above tree, the value of the object argument of muita and the
head of the LINK ed structure must be the same. The rule having been applied,
the head formula must then be supplied by some following expression of type
e provided by the immediately following natural language string; so, unlike the
parsing of successive noun phrases in the Japanese string, at this point in the
interpretation process, there is a procedural location where the structural role
of the immediately subsequent part of the string is strictly determined. This
rule thus provides the basis for the interpretation of the head-nal structure,
with ringo (= `apple') immediately following the verb muita (= `peel').
The next step in the sequence of actions is the processing of ringo, with its
lexical entry:
ringo
IF f?T y(t)g
THEN make(hDi); go(hDi); put(?T y(e))
DO()
ELSE IF f?T y(e)g
THEN DO()
ELSE ABORT
where DO() = put(F o(U); T y(e)); make(h"0 ih#1i);
go(h"0 #1 i); put(F o(Ringo); T y (e ! cn); [#]?);
go(h"1 i); put(?T y (cn)); make(h"0 i); go(h"0 i); put(?T y (e));
make(h#1 i); go(h#1 i); put(F o(P (; P )); T y (cn ! e));
go(h"1 i)
The sequence of actions dened for ringo has rst the general condition of
requiring T y(t) (which provides the cases we have already seen, where ringo
may project and annotate an unxed constituent of type e); but it should also
be prepared for the structurally specic location immediately following a verb.
In the latter case the noun must project the head of the previous relative clause.
That is, following a verb, the structural role of a noun is as fully determined
as it is in a language like English. If the condition of this word is satised, it
creates a tree structure annotating the current node (requiring type e) with the
nominal variable, but embeds that in a larger structure. The resulting structure
for the interpretation of the object noun phrase in (4.67) (with the additional
restrictor predicate added from the LINK ed structure) has the form displayed
in gure 4.9.
The lexical actions projected by the common noun ringo immediately after
the construction of the LINK transition identify the nominal variable with the
type e required at the type e node resulting from the transition: this creates an
interpretation as a restrictive relative. To the top node of the noun projection
the accusative case marker o then adds the requirement ?h"0iT y(e ! t), after
Relatives: Towards a Dynamic Typology 137
fF o(; (x; Ringo(x));
T y (e)g
F o(U) F o(Ringo)
F o(Muita( U)(Hiroto))
F o(Hiroto) Fo(Muita(U))
F o(U) F o(Muita)
From the dynamic perspective, the interest of head-nal relatives is two-
fold. First, the tree growth process has to allow upward development from
some node in linking one tree with another, so development is not exclusively
top-down. The second distinguishing property in the sequential projection of
these structures is the need to allow a process of node construction which allows
a node to be introduced which bears the weak tree relation D to the root. This
is not an option that we have allowed hitherto; and it re
ects a genuine cross-
language dierence as to the degree to which the interpretive process is goal-
directed and driven by top-down expectations of structure, or, to the contrary,
tolerant of structure being built only as triggered by lexical expressions. The use
of the hDi operator to project sequences of nodes otherwise unrelated re
ects
directly the way in which a language such as Japanese presents a
at sequence
of NPs while nevertheless projecting a structure for interpretation with familiar
Relatives: Towards a Dynamic Typology 139
congurational properties. Here we merely note that one possible transition
amongst the disjunction of possibilities at the opening steps within the construal
of a clausal sequence is to allow the building of nodes which are only weakly
related to the node that constitutes the point of departure. It should be noted
that this is a possibility we make available here only from the root, and not
more generally.
Apart from these two distinguishing features, the process of inducing struc-
ture remains locally top-down, the verbs invariably building propositional struc-
ture from a trigger of type t with a set of actions that annotates all the nodes
within that structure. It is then34 subsequent operations of Merge or Substitution
which yield the desired result.
fF o(;U, Ronbun(U)),
T y (e)g
38 It should be noted that there appear to be pragmatic restrictions on what can and what
cannot be expressed in the form of a head-internal relative, but we leave these on one side,
for further investigation.
39 Notice how examples such as (4.69) provide further demonstration that quantied expres-
sions may give rise to non-restrictive interpretations. See n.12.
142 Linked Tree Structures
40 There are no such head-internal relatives in Chinese, which at rst glance is surprising.
But, in Chinese, the same form of interpretation is achieved by the head-nal structures. In
Chinese, the sequence of words which immediately follows the nal relativizing element de in
head-nal structures is the full noun phrase sequence of quantier-classier-nominal:
(i) ai ta muqin de mei ge ren
loves his mother rel every classifier man
`every man who loves his mother'
With de projecting both a meta-variable for one argument of the verb and the LINK transition
from some dominating node of type t, it is what follows de that will determine the term
with which the meta-variable will have to be identied; and in this type of structure it is
the term projected from the quantier-classier-nominal sequence which projects the shared
term, not just a nominal meta-variable. As we would expect, this gives rise to non-restrictive
interpretations, again notwithstanding the quantifying determiner.
41 No is also used as a genitive marker, which, as we are about to argue, also induces a LINK
relation.
Relatives: Towards a Dynamic Typology 143
expression (see Keenan 1985 for a broad typological survey), a fact which ts
our analysis very well.42
It might seem that the framework is over-liberal in allowing so much vari-
ation, with lexically induced and rule-based LINK transitions, both cross-
linguistically and even within a single language. It should be pointed out, how-
ever, that each parameter for variation is an independently motivated property
of the left-to-right form of tree growth { there is the distinction between an
annotation on a node and a requirement; the distinction between whether a
requirement is to be locally or globally satised; and interaction between com-
putational, lexical and pragmatic actions. The type of tree decoration and the
incremental form of goal-directed tree development are also common to all ex-
planations. Nevertheless, within this theoretical space, variation is indeed freely
available both cross-linguistically and within a single language. This provides a
42 Kayne's (1994) analysis shares with the present one a concern to respect considerations
of linear order yet give a unied analysis of relative clauses. According to Kayne's analysis,
non-restrictive, restrictive, head-initial, head-nal and head-internal relatives are all assigned
a base structure in which the Determiner has as complement a CP containing the full wh
expression together with the common noun plus any accompanying adjectives as a DP in the
clause-internal position. This full wh DP is moved overtly to the Spec CP position, and the
nominal associated with the wh determiner is moved over the wh within this internal DP to
Spec DP position to yield:
[D0 the [CP [DP bookj [D0 which tj ]i [C0 [IP I read ti ]]]]]
Non-restrictive relatives involve further movement, at LF, to yield a conguration in which
the IP is moved further up to the Spec DP conguration { providing a basis for its construal
independent of the DP of the common noun to which it is adjoined. Head-nal relatives have a
similar underlying structure, but require overt movement of the entire IP to Spec DP position
(or possibly to some node dominated by the DP node, but dominating the NP node within
which the relative sequence originated). Head-internal relatives are then alike, except that,
with movement construed as copy and delete, it is the copy of the NP which is deleted, leaving
the full NP expression within the relative sequence, contained at the highest level within the
Spec DP position.
Aside from the richness of movement processes, the analysis is not problem-free. It requires
that the nominal remains part of the CP structure, the entire nominal-plus-relative sequence
taken as complement to the determiner. This suggests, wrongly, that noun-plus-relative se-
quences should be pronominalized, like CPs, by the pronoun it :
(i) That book which I left on the table is not as nice as this *it/one.
Pied-piping cases are problematic (with no copy of the noun relative to which the adjoined
relative is an adjunct), particularly non-restrictive ones, where the adjoined wh expression
may contain a non-identical nominal:
(ii) The book whose author I disapprove of has just been published.
(iii) That book, which horrendous piece of writing should never have been accepted for
publication, is making enormous amounts of money.
Furthermore, with only a single occurrence of the nominal, within the CP structure, there
is no basis for characterizing how a noun which is head of some relative can be understood
independently of the relative adjoined to it:
(iv) The hostages who were now arriving had been very badly treated.
(v) The hostages who had been very badly treated were now arriving.
In order to project such an interpretation, one might posit movement at LF, but this would
only characterize non-restrictive forms of construal (not (iv){(v)). There is, furthermore, no
basis for explaining the distinction between the interpretation of head-nal and head-internal
relatives (see Hoshi 1995), or more generally, the correlation between anaphora and relative
clause construal.
144 Linked Tree Structures
new perspective on language variation { that any variation licensed within the
framework may occur in any one language. This turns out to be a bonus, as we
will now see.
4.4 Genitive Constructions as LINK Structures
Nothing precludes the construction of a pair of linked trees with the LINK ed
tree constructed prior to the projection of the head. Such a dierent direc-
tionality might seem a prototypical basis for typological dierences between
languages according as the language is head-initial (e.g. English) or head-nal
(e.g. Japanese). But, as we shall now see, head-nal structures are available in
English as lexically triggered.
In English genitives, a modifying structure of arbitrary complexity precedes
the structure modied, and there is, in consequence, an essential process of
upward tree growth:
(4.71) John's mother's friend died.
Informally, the initial task in parsing such a complex noun phrase sequence,
is the parsing of a full noun phrase, John, as though subject. As the string
continues, it becomes apparent that the formula projected from John is associ-
ated with an increasing level of embedding in the subject node. The sequence
of actions involved in parsing the subject noun phrase is one of constructing a
sequence of nested structures { in other words, remarkably redolent of Japanese
pre-modifying structures, albeit in a language in which the subject and predicate
nodes are constructed by the rules Introduction and Prediction.
In this kind of construction, the Adjunction rule becomes essential, for the
node created to accommodate the subject information is a daughter of the top
node (requiring type t). The genitive construction must apparently now `inter-
polate' nodes between the top and the putative subject node, and this is an
operation which would not preserve tree structure. So for such structures, we
use Adjunction to create an intermediate node:
fX; ?T y(e); }g
fX; : : : ; ?T y(e)g; fhU iX; ?T y(e); }g
Adjunction, as dened by this rule, allows the creation of an alternative separate
node, which can then be embedded in a larger structure. From this newly
introduced node, the construction of a nested sequence of type e nodes projected
by a genitive NP such as John's friend is then induced by the suÆx 's, which
separates the possessor (John) from the possessee (friend) in the complex NP.
The actions projected by this suÆx are rst to embed the possessor in a tree
with top node of type t and subsequently link the root node of this tree to a head
node in which a term for the possessee is embedded. The eect of the possessive
suÆx 's in the interpretation of John's mother can be seen as falling into two
parts. The rst part, starting from the projection of John, nests that node
Genitive Constructions as LINK Structures 145
within a containing tree involving a binary P OSS relation with the projection
of John and a possessee variable as the two arguments:43
fX; F o(John); T y(e); }g
fF o(P OSS (U)(John)); T y(t); }g
43 P OSS is a relation whose value is constructed from some contextually provided value.
We leave it here in the form of P OSS for simplicity.
146 Linked Tree Structures
f?T y(t)g
(4.73) *The children who Sue said's mother was lazy were often late for
school.
Finally, we also have a basis for characterizing the observed dierence in inter-
pretation between (4.74) and (4.75):
(4.74) John's picture
(4.75) the picture of John
(4.75) can only be construed as the picture which is a portrayal of John; (4.74)
allows any arbitrary relation between John and the picture. On the analysis
suggested here, only (4.74) is taken to lead to the construction of a relation
P OSS between the two terms John and (; x; P OSS (John; x)). The preposi-
tional complement of the common noun picture is, however, taken to project a
second argument to a two-place predicate projected by the common noun, and
must be correspondingly interpreted.
This analysis of the genitive as a form of relative is far from new. Indeed,
it is novel only by virtue of the dynamic perspective within which it is set.48 A
theoretical consequence of postulating this analysis for a construction in English
is that we allow that even structures not generally thought to be available in
a head-initial language may nevertheless be manipulated in lexically specic
environments. We return to this in chapter 8, merely noting here that the
concept of cross-language variation as a parametric switch xed in language
acquisition is not a natural consequence of the dynamic perspective articulated
in this framework (see Chomsky 1981, Manzini and Wexler 1987).
4.5 Summary
In this chapter we have developed an account of relative clauses which abandons
the orthodox variable binding operator account of relatives. The account de-
pends on modelling the way in which information is built up through a left{right
dynamics, re
ecting the result achieved by operator variable binding in other
frameworks through an encoded copy device which is an essential intermediate
step in this build-up of interpretation. The major dierence between this and all
48 Further evidence for this form of analysis in other languages is given by Baker (1996),
Kayne (1994). It should be noted that this analysis has no explanation for the occurrence of
anaphors within `picture-noun-phrase' constructions such as John's picture of himself. How-
ever following Pollard and Sag (1994) and Dalrymple (1993), in the face of examples such as
(i){(iii), we assume that the locality restriction constraining anaphor construal is not denable
over tree structure, but is, rather, more indirectly, a trigger for inference of a form yielding
some formula (; ):
(i) John's publicity campaign ensured that there were pictures of himself on every bill-
board.
(ii) John was on the point of leaving the show on the grounds that Mary's portrayal of
himself had been sharper than he had expected.
(iii) John walked in. Pictures of himself were all over the
oor.
Summary 149
other approaches is the broadening of the basis for explanation. The emphasis
is no longer merely on construction of a structure re
ecting semantic interpre-
tation, but on the sequence of steps needed to establish a route from the basic
requirement for a proposition as interpretation of the natural language string to
the nal result. At every step, emphasis has been placed on what information
is available at that step of the process.
The vindication of this approach lies in its two-fold success. First, it has
made possible cross-linguistic accounts of relative construal, while sustaining
a unitary and structural account of how pronouns are construed. Secondly, it
has led to accounts of head-initial, head-nal and head-internal relative struc-
tures which both integrate the various sub-types of relatives within a unifying
typology, and yet re
ect their dierences. In all types of case, the various
sub-types and their specic instantiations in dierent languages have emerged
from general principles of how tree descriptions are monotonically built up. On
the one hand, resumptive pronouns occurring in both relative clause and topic
structures are explained without having to postulate any construction-specic
sub-type of pronominal, merely in virtue of there being two sorts of decoration
on a node { annotations and requirements { the latter inducing subsequent tree
development. On the other hand, the dierence between head-initial and head-
nal relatives is seen to emerge largely as a consequence of the dierence in
ordering; and the head-internal relative constructions, so problematic for other
frameworks, emerge as a natural part of the overall analysis, with the common
term of the two resulting structures licensed to occur in the LINK ed structure
in virtue of the linear order. And a characterization of genitive constructions ts
nicely within the overall family of explanations. Underlying each explanation
has been the concept of tree growth, with descriptions of a tree being built up
monotonically from some starting point through to some nal completely anno-
tated linked tree structure, with no requirements outstanding. The core of the
explanation is thus the dynamics of how structured information is incrementally
developed.
5
Wh Questions: A General
Perspective
5.1 Introduction
We have so far simply taken for granted that the characterization of gap-
containing relative clauses and left-dislocation in terms of an initially unxed
node will extend to the full range of phenomena analysed in other frameworks
by an operator-gap binding mechanism. Accordingly, wh expressions in chap-
ter 4 were dened as projecting a meta-variable WH that, if the wh expression
is initial in a clausal sequence, annotates an unxed node. This analysis of
wh-initial structures
ies in the face of the orthodox assumption that wh ex-
pressions are variable binding operators, located in a structure at some Spec
CP position, from which they bind a discrete position (see Chomsky 1981 and
many references thereafter). Even formalisms which posit no gap position as
such, none-the-less characterize wh expressions as variable binding operators in
the sense that a process is dened at the node in the tree they inhabit, which
discharges a feature/assumption annotating a discrete node somewhere lower in
the tree (Steedman 1996, Sag and Fodor 1994, Jacobson 1996, Morrill 1994).1
In this chapter we take up our alternative analysis of wh structures in more
detail by looking at dierent forms of wh questions. First, we shall set out some
semantic problems facing accounts of wh questions in terms of a variable binding
operator. Then we shall show how the framework articulated here not only
captures wh-initial expressions as long-distance dependency phenomena, but
also straightforwardly characterizes wh-in-situ expressions. We shall go on to
demonstrate the fruitfulness of this model by taking up the problematic expletive
1 Analyses within HPSG dier according to whether the SLASH feature proposed con-
stitutes a distinct node in the tree or is projected by a verb, but the assumption that a
left-peripheral wh expression in a clause inhabits a node re
ecting its left-peripheral position
is assumed by all (see Pollard and Sag 1994, Sag and Fodor 1994, Sag 1997, Bouma and
Malouf 2000).
150
The Semantic Diversity of wh Questions 151
wh constructions, and show that these expressions can be analysed as projecting
as a requirement what the non-expletive form projects as an annotation, an
analysis which enables us to explain the process of so-called `partial movement'
which is a characteristic feature of expletive structures. So we shall see that the
dynamics of tree growth which yields a typology for dierent forms of relative
clauses across languages also provides a basis in the apparently rather dierent
area of wh question structures.
5.2 The Semantic Diversity of wh Questions
Generally, work on the semantics of wh questions has started from the premise
that assignment of semantic objects to interrogative strings in a compositional
way is a minimum requirement for any account. It is therefore assumed that an
interrogative, like an indicative, should be assigned some form of denotational
content, the standard account being that the content of a question is given by
partitioning the universe of discourse to pick out some set of possible answers
(see Groenendijk and Stokhof 1997 for an overview of attempts to provide a
semantics for questions and a defence of this form of analysis). The slogan is
`To know the meaning of a question is to know what counts as an answer'.
Within this general set of assumptions, there are two major alternative char-
acterizations of wh questions: as an abstraction operator denoting a union of
possible partitions (see Groenendijk and Stokhof 1989, Higginbotham and May
1981, Higginbotham 1996), and as a form of existential quantication denoting
some subset of possible answers (see Karttunen 1977). The problem which these
accounts face is that neither of them is suÆcient to cover all the data, giving
rise to a phenomenon of multiple analyses for a single type of expression, much
as with anaphoric expressions. The question in (5.1), for example, seems to ask
for a listing of all places where poison was put down, but the question posed in
uttering (5.2) does not require any such exhaustive listing:
(5.1) Where did the ratman put down the poison? The sitting-room, the
kitchen and the larder.
(5.2) Where can one buy an Italian newspaper? At the station.
Evidence that this requires an analysis positing dierent types of wh expres-
sions is taken to be provided by the corresponding embedded questions (see
Groenendijk and Stokhof 1997), whose interpretation is presumed to consti-
tute some (exhaustive or non-exhaustive) set of possible answers.2 For, despite
structural equivalence between individual embedded questions, they may nev-
ertheless be interpreted in systematically dierent ways. The truth of (5.3)
depends on John knowing every place where the ratman put down poison: we
would hardly agree that John knew where the ratman put down poison if he
did not know for each room in the house whether or not poison had been put
2 For extensive argumentation against this assumption, see Ginzburg 1992, 1995.
152 Wh Questions: A General Perspective
down there. (5.4), on the other hand, asserts only the weaker claim that John
knows at least one place where an Italian newspaper can be bought:
(5.3) John knows where the ratman put down the poison.
(5.4) John knows where one can buy an Italian newspaper.
Given a standard principle of compositionality and the assumption that the
meaning of expressions is given by their denotational content, such data thus
appear to indicate ambiguity in the wh expressions themselves. Both types of
analysis of wh expressions nevertheless assume that the scope of the wh operator
is indicated by the surface position of the wh expression, c-commanding the
clause to which it is adjoined.
The dierent forms of analysis required for wh questions are not exhausted
by the apparent potential for universal or existential construal of the wh ex-
pression. Evidence has been put forward by Reinhart (1997) that wh-in-situ
constructions require a discrete form of analysis to capture their supposed con-
strual externally to their containing context:
(5.5) Which lady wants to read which book?
(5.6) Who will be annoyed if we invite which philosopher?
As she points out, the in-situ form must not be construed from the position
internal to the conditional internal to (5.6) (on the assumption that the condi-
tional corresponds to material implication) if we are to debar answers to (5.6)
such as (5.7), for the falsity of Donald Duck being a philosopher will have the
unfortunate consequence of satisfying the structure in (5.6) on the assumption
that the wh form is some type of indenite contained within the antecedent of
the conditional:
(5.7) John will be annoyed if we invite Donald Duck.
The choice function analysis which she proposes has the advantage of being able
to characterize wide scope construal of the wh expression vis-a-vis its containing
context while nevertheless being understood as taking narrow scope of the in-
situ wh expression vis-a-vis the other wh expression (see Reinhart 1997, Winter
1997), though such an account of wh-in-situ constructions, distinguishing their
denotational content from wh-initial expressions, has the eect of characterizing
(5.8) in two dierent ways, depending on whether the rst wh expression is
analysed as a regular wh operator or as a wh-in-situ expression:3
(5.8) Which woman read which book?
Despite this addition to the types of wh expression, the list is still not complete.
The assumption that the semantics of a question is given by its possible an-
swers, together with the assumption that a wh expression is a variable binding
3 Reinhart (1997) only considers the diÆculty which wh-in-situ constructions pose for the
Karttunen-style analysis of wh expressions as a form of indenite.
The Semantic Diversity of wh Questions 153
operator whose scope is given by its structural position, gives rise to yet more
ambiguity. It appears that a discrete type of so-called functional questions must
also be dened for wh-initial questions, for in questions such as (5.9) { despite
the structural position of the wh expression c-commanding the quantied ex-
pression { the form of answer demands a construal in which it is the universally
quantied expression which contains everything else within its scope, and not
the questioned element (see Engdahl 1986, Chierchia 1992):
(5.9) Who did everyone recommend? Their best student.
p9f [p = 8x(recommend(x; f (x)))]
`What is the function that pairs people with those they recommend?'
Note that in this case it is the assumption that an initial wh expression marks
its scope relative to the surface string which requires the analysis of wh expres-
sions as a function that picks out an individual relative to some antecedently
articulated quantier. It should be noted that this is not a locally denable
dependency between a pair of quantifying expressions within the same clause.
It would appear, rather, that a wh expression may be construed as having ar-
bitrarily narrower scope than some quantier that it c-commands:
(5.10) What is the Union insisting that almost every British farmer should
get rid of? At least 1,000 cattle.
We know that this phenomenon is not due to the wide scope potential of the
lower quantier in a case like this, as a quantier such as almost every may not
be construed as taking scope beyond the clause in which it is contained { (5.11)
has no such wide scope interpretation of the embedded subject:
(5.11) Most British farmers are complaining that almost every country of the
EU fails to appreciate the problem.
There is yet a further type of question, a so-called choice interpretation, posited
for a pair of disjunctive questions. In these, the answer cannot be construed as
the union of the two answers (as the partition theory of Higginbotham 1996 or
Groenendijk and Stokhof 1989 would lead one to expect), and so the question
re
ecting such answers has to be analysed as something other than a regular
disjunction:
(5.12) Where's your father? Or where's your mother?
This problem also arises even within a single sentence, such as:4
(5.13) Who likes Bill or Jill?
Finally, no single one of these suggested denitions of wh expressions provides
a basis for characterizing questions where the set of possible answers is unclear
or unspeciable (open questions):
4 Thanks to Jonathan Ginzburg for pointing this out to us. See Ginzburg and Sag (1999)
for extensive discussion of the problems which disjunction poses for Groenendijk and Stokhof's
account.
154 Wh Questions: A General Perspective
(5.14) What are questions?
(5.15) What is Buddhism?
(5.16) What causes a sun's death?
At best, the characterization of a question as the set of possible answers whose
answers are unclear, vague or unspeciable is that the question is correspond-
ingly unclear, vague or unspeciable. But this contradicts common sense: it
does not follow that a question each of whose answers is unclear or vague or
unspeciable is necessarily an unclear, vague or unspeciable question.
The two assumptions on which such analyses of questions are based thus
lead to a multiplicity of types of question. As Groenendijk and Stokhof observe
(1997), the situation is redolent of anaphora, with no single answer to the ques-
tion `What is the meaning of a wh question'. In all cases, the answer is merely
that the meaning of a wh question is one of a number of things, recovery of
which, in context, depends on recovering the type of answer the speaker has
in mind. Not surprisingly, given this unsatisfactory conclusion, as in the lit-
erature on anaphora, there have been various attempts to reduce the observed
heterogeneity, none entirely successful.
5.2.1 Scopal Properties of wh Expressions
There is a further set of puzzles as to what concept of scope should be attributed
to wh expressions. Groenendijk and Stokhof and many others following them
assume that the position of the wh expression in the string determines its scope
as a variable binding operator. However, this poses a number of problems. First,
with the so-called functional questions apparently demonstrating the construal
of the wh form as having narrower scope than other quantifying expressions
in the string, it might seem that there is ambiguity according to whether the
position of the wh expression dictates its scope relative to other expressions.
Second, in addition to either of these possibilities, there is, in some languages,
a discrete form of wh, the expletive wh, which is said to extend the scope of the
wh to that which is wider than would otherwise be projected from its surface
position (see Pafel 1996, Johnson and Lappin 1999, Pollard and Yoo 1998, etc.).
Thus, in German, (5.17) and (5.18) are synonymous. In both cases, so the
explanation goes, the scope of the wh operator is that of the highest clause:
(5.17) Was glaubst du, mit wem Jakob gesprochen hat? German
What think you with whom Jacob spoken has
`With whom do you think Jacob was talking?'
(5.18) Mit wem glaubst du, dass Jakob gesprochen hat?
With whom think you that Jacob spoken has
`With whom do you think Jacob was talking?'
The eect of the expletive was is reported by McDaniel (1989) to indicate in
(5.17) that the lower mit wem is to be construed as taking wide scope with
The Semantic Diversity of wh Questions 155
respect to the higher verb glaubst in the same manner as (5.18). One stan-
dard analysis of these constructions is accordingly that the expletive is a scope
marker, exceptionally extending the scope of the lower wh expression to that of
the clause marked by the expletive (more accurately, the highest of a sequence
of such expletives):
(5.19) Was glaubst du, was Hans meint, mit wem Jakob
What think you what Hans said with whom Jacob
gesprochen hat?
talking has
`With whom do you think Hans said Jacob was talking?'
However, despite this supposed scope-marking property of the expletive form,
when such question forms are combined with other quantifying expressions, the
eect is to strongly encourage readings in which the wh expression is interpreted
as taking narrower scope than the quantier between the expletive and full wh
form (see Pafel 1996 for arguments to this eect):
(5.20) Wo glaubt jeder, dass sie gerne leben wurde
Where thinks everyone that she live like would
`For which place y, for each x, x think she would like to live in y?'
`For each x, for which place y, x think she would like to live in y?'
(5.21) Was glaubt jeder wo sie gerne leben wurde
What thinks everyone where she live like would
`For each person x, for which place y, x think she would like to live in
y?'
`For which place y, for each x, x think she would like to live in y?'
A dierent puzzle is posed by wh-in-situ expressions. As some form of a quan-
tifying expression, these might in principle be expected to be interpreted as
falling within the scope of other quantifying expressions. But, in fact, they are
only sensitive to some other wh expression in the sentence, and not to any other
c-commanding quantier. Thus (5.22) is ambiguous, but only in two ways, de-
pendent on whether the wh-in-situ form is construed relative to the embedded
wh expression or to the higher wh expression. Hence the two possible types of
answer (a) and (b) to (5.22):
(5.22) Which teachers told every student who should concentrate on which
19th century author?
(a) Bill told every student who should concentrate on Wordsworth;
Sue told every student who should concentrate on Henry James.
(b) Bill told every student who should concentrate on which author;
and so did Sue.
We are thus faced with data for which the various analyses do not seem to
add up to a consistent overall analysis. The position of the wh expression has
156 Wh Questions: A General Perspective
been taken in the main to indicate its scope over other expressions in the string;
although sometimes (in functional interpretations) it does not. The scope eects
of its position are apparently overridden in two ways { either by some process
construing the wh expression as dependent on some subsequent expression in
the string, or by the presence of a scope-marking expletive apparently extending
the scope of the wh expression so that it has scope over expressions before it in
the string. In this latter case extra problems arise, because that wh expression,
notwithstanding the presence of the expletive, is in some examples construed as
having a scope considerably narrower than that marked by the expletive. And
nally, unlike all the other wh forms, an in-situ wh doesn't seem to interact
with any quantier other than other wh expressions. A single base from which
to explain these various wh forms might seem to be a goal that has to be
abandoned.
These analyses of wh expressions have been, without exception, grounded on
syntactic explanations given in terms of tree congurations decorated by words
of the string (or at least abstract characterizations of those words). Against this
background, the assumption that a wh expression is a variable binding operator
has been taken to be uncontentious, its binding properties dened in terms of
the tree relation between the node it decorates and nodes decorated by other
words. We formulate an alternative analysis not based on these assumptions,
in which a wh expression underdetermines its contribution to the interpretation
of the whole. According to our view, attributing scope to a wh expression is an
attribution of properties of some enriched form { the answer { to the incomplete
structure { the question; and so it does not characterize the properties of that
input structure independently of its available enrichments. What we propose is
that wh expressions project a distinguished meta-variable, WH, which, if initial
in a clausal sequence, is projected as an annotation of an unfixed node in the
logical form under construction. As we shall see, the dynamics of the growing
of tree structures and the progressive xing of tree node positions will, as in the
case of relative clauses, enable us to characterize a supposedly heterogeneous
set of data in a revealing and natural way.
5.3 Wh -initial vs wh -in-situ Structures
The rst observation about the projection of some WH meta-variable, from the
perspective of the present framework, is that the annotation of a node with a
meta-variable is independent of whether that node is or is not xed: in either
case the nal result will be a tree whose root node is annotated with a formula
containing a WH meta-variable. Upon this analysis then, a pair of questions
such as (5.23) and (5.24), whose primary dierence is that one has the wh ex-
pression at a left-dislocated position and the other at an in-situ position, will
Wh-initial vs wh-in-situ Structures 157
give rise to the same formula at the root node of the resulting tree (displayed
schematically in gure 5.1):5
(5.23) Who did Bill see?
(5.24) Bill saw who?
fT y(t);
F o(See(WH)(Bill)); }g
5 As in other tree displays, we suppress all information except that which is essential to
the point being made. This analysis of wh questions is similar to that of Ginzburg and
Sag (1999), who argue that a wh expression is not a generalized quantier but a specialized
variable { in their terms, a restricted index. However, in construing the concept of index and
its associated uniqueness presupposition semantically, they are forced to postulate ambiguity
as between functional construals of wh questions and what they take otherwise to be wide
scope construals for all wh expressions in questions, and also as between dierent construals
of which, as between what they call n-jective construals and independent construals (both of
which they take to involve no scope dependence). They grant that if they adopted the stance
that such uniqueness presuppositions were only induced as a general pragmatic consequence,
as argued by Engdahl (1986) { a position with which we would concur despite Ginzburg and
Sag's arguments to the contrary { a unitary account would be available. As they themselves
observe, reporting the observation of Engdahl (1986), such uniqueness presuppositions are in
any case not sustained in multiple wh questions.
158 Wh Questions: A General Perspective
associated with verbs such as wonder that the wh expression identifying the
complement clause as a question must occur initially:6
(5.25) John wondered who Bill saw.
(5.26) *John wondered Bill saw who.
As we shall see shortly, the feature Q is also critical for determining the distri-
bution of wh forms in the presence of the expletive German wh form.
The meta-variable WH projected by the wh expression can remain unre-
placed in a resulting formula.7 Furthermore, the wh expression itself will not
have its scope xed with respect to other terms in the resulting formula, so a
quantied term contained within a wh question may fail to have its requirement
of an associated scope statement fullled. In this sense wh questions are essen-
tially incomplete { a full semantic evaluation cannot be dened until an answer
is provided. As we shall see in due course, this leads to a dierent account
of scope in wh questions. Before reaching that point, however, we explore the
dierent kinds of analyses which the assumption that wh expressions project a
meta-variable provides.
5.3.1 Wh -in-situ Structures
In Germanic and Romance languages, the so-called wh-in-situ expressions are
standardly reported to be peripheral { normally, though not invariably, asso-
ciated either with an echoic use (characteristically associated with a marked
intonation) or as a second member of multiple wh expressions:
(5.27) John saw who?
(5.28) Who do you think saw what?
However, as pointed out by Ginzburg (1992, 1995), they are associated with a
number of dierent uses:
(5.29) When you get to the oÆce, your rst action will be to do what?
(5.30) Having arrived there, you did what exactly?
In other languages, they form the regular form of question, as in Chinese:
6 We might also posit an additional feature QUE for characterizing all questions, both
yes-no and wh. Each of these features is introduced as a DU formula of the form Cat(Q),
Cat(QUE), etc., picking out a dierent set of structures of type t. It is the feature QUE that
we might use to dene the conditions inducing auxiliary subject inversion in root clauses in
English, a phenomenon which is arguably a re
ex of a verb-adjunction process carried over in
a more specialized form from German. In German, all verbs may be introduced at a locally
unxed node, but in English, where this process has atrophied, the option of building such a
node is an action specic to auxiliary verbs (this context for auxiliary occurrence was ignored
in the lexical specication of auxiliary verbs in chapter 3). See sections 3.2.2, and 5.4.2.1.
7 We shall consistently use `WH' to indicate the meta-variable assigned as the F ormula
value, and `wh' to indicate the natural language expression which projects such a meta-
variable.
Wh-initial vs wh-in-situ Structures 159
(5.31) Ta shuo shenme [Chinese]
He say what
`What did he say?'
In most languages, in-situ uses are not restricted to particular kinds of structure.
A wh expression is a noun phrase, and can occur anywhere that noun phrases
can { in relative clauses, sentential subjects, etc. In other words, wh-in-situ
expressions are generally not subject to any of the island restrictions familiar
from long-distance dependency:8
(5.32) The journalist that leaked which document to the press became
famous overnight?
(5.33) That he had missed which movie upset him ?
(5.34) Ni mai-le [[shei xie] de shu] [Chinese]
You buy-asp who write rel book
`Who is the x such that you bought books that x wrote?'
(5.35) Lisi da-le shei shi ni hen bu gaoxing
Lisi hit-asp who make you very not happy
`That Lisi had hit who made you unhappy?'
However in some languages, the occurrence of wh expressions in situ is subject
to more stringent restrictions than in wh -initial structures. For example, in
Iraqi Arabic, wh expressions may normally occur in subordinate clauses only
if the clause is non-nite. If the embedded clause is nite, only the wh-initial
form is possible:
(5.36) Mona raadat [tijbir Su'ad tisa'ad meno] [Iraqi Arabic]
Mona wanted to force Suad to help who
`Who did Mona want to force Suad to help?'
(5.37) meno Mona raadat [tijbir Su'ad tisa'ad]
who Mona wanted to force Suad to help
`Who did Mona want to force Suad to help?'
(5.38) *Mona tsawwarat [Ali ishtara sheno]
Mona thought Ali bought what
(Intended: `What did Mona think that Ali bought?')
(5.39) shenoi tsawwarit Mona [Ali ishtara ti]
What thought Mona Ali bought
`What did Mona think Ali bought?'
8 The one possible exception is why-type adjuncts which are reported not to be able to
occur inside relative clauses (Huang 1982).
160 Wh Questions: A General Perspective
In Iraqi Arabic, wh expressions, when in situ, are thus licensed only in environ-
ments in which antecedent anaphor relations are.
Systematically across languages, this pattern of wh -in-situ structures carries
over to multiple wh structures: these display a combination of wh -initial prop-
erties and wh -in-situ properties. In English the initial wh cannot be interpreted
with the supposed gap inside a relative clause (an unxed node cannot nd its
position across a LINK relation), whereas the second wh may occur anywhere
{ for example, inside a relative clause:
(5.40) Who do you think should review which book?
(5.41) *Whoi did the journalist leak the document in which Sue had
criticized ei to which press?
(5.42) *Which documenti did the journalist that leaked ei to the press resign?
(5.43) Who reported the journalist that leaked which document to the press?
In Iraqi Arabic, to the contrary, the second of a pair of wh expressions may occur
in an embedded structure projecting multiple wh-question interpretations only
if the clausal structure in which it is contained is non-nite:
(5.44) sheno ishtara Ali [minshaan yenti li-meno] [Iraqi Arabic]
What bought Ali in order to give to whom
`What did Ali buy to give to whom?'
(5.45) *meno tsawwarit [Ali xaraj weyya meno]?
Who thought Ali left with whom
`Who thought that Ali left with whom?'
This array of properties is problematic for analyses of wh-in-situ forms in which
these undergo the same movement as wh-initial forms but covertly (see Huang
1982 among others), as the constraints on this movement diverge from familiar
subjacency restrictions: they are either more stringent, as in Iraqi Arabic, or
less so, as in English, Chinese etc.9 Multiple wh questions, as Simpson (2000)
9 These and other diÆculties are presented in detail in Simpson (2000), who argues that
features must be checked from their in-situ position, hence in congurations other than a
local Spec Head conguration. However, the checking of features for wh expressions from
an in-situ position poses its own problems, because this raises the question as to why move-
ment should ever take place. Simpson (2000) argues that movement to a Spec CP position of
the wh expression is obligatory in order to `disambiguate' an otherwise `crucially underdeter-
mined' complementizer, but this analysis hovers unsatisfactorily between underspecication
and ambiguity. If the complementizer in question is a `general polarity-oriented head' and is
underspecied with respect to any more specic value, then it will lack any feature particular
to wh forms to license wh-in-situ forms. Ambiguity, on the other hand, is by denition re-
ected in the grammar in the form of discrete items, and on this analysis, each such particular
type of `polarity-oriented' complementizer will project the particular feature needing to be
checked; but then no movement will be necessary. Despite Simpson's informal statements,
conventional assumptions about grammar-parser relations do not permit problems posed by
the need to disambiguate a string as between one of several independent structures to be
re
ected in the grammar formalism, for this is not a phenomenon intrinsic to any one of the
structures posited.
Wh-initial vs wh-in-situ Structures 161
demonstrates, provide additional problems for Minimalist assumptions, because
the feature +WH on the wh expression is a strong feature to be checked by
adjacency to a node bearing the feature Q triggering movement from some in-
situ position; but once checked, the whole structure is taken to be licensed for
the wh feature, allowing multiple occurrences of a given wh expression. The
problem with Iraqi Arabic multiple wh structures is that the restrictions on the
wh-in-situ form are carried over to these structures, even though the prohibited
environments for the in-situ form are the very environments in which movement
must take place.
5.3.2 Wh -in-situ from a Dynamic Perspective
In the present analysis, the10in-situ forms are simply the annotation by a wh
expression of a xed node. The lexical characterization of what in English
re
ects the two types of structure:
what IF f?T y(e)g
THEN IF fh" i?T y(t); ?9xT n(x)g
THEN put(F o(WH); T y(e); [#]?);
go(h" i);
put(Cat(Q))
ELSE put(F o(WH); T y (e); [#]?)
ELSE ABORT
The condition stipulates that should the node to be annotated be unxed, then
the resulting action is to add the DU formulae F o(WH), T y(e) and annotate
the dominating node with Cat(Q), and that otherwise, it simply adds the DU
formulae F o(WH),T y(e).11 As indicated above, the feature Q is a classicatory
device for picking out wh-initial questions without semantic consequence.12
In the in-situ cases, no such feature Q is added: the only property distin-
guishing wh-in-situ structures from non-wh question structures is the meta-
variable WH itself. It follows immediately that there will be no restrictions
precluding a wh-in-situ expression from occurring in relative clauses or other
environments which do not license long-distance dependencies, as there is, on
this account, no long-distance dependency to establish. Wh -in-situ expressions
also occur in non-restrictive as well as restrictive relative clauses (contra any
analysis which analyses wh-in-situ expressions as subject to covert movement
at LF):
10 We follow Ginzburg (1992) in assuming that in-situ wh forms should be licensed in Eng-
lish as freely available. See Ginzburg 1992 for detailed discussion of in-situ wh questions,
and evidence demonstrating that English wh-in-situ phenomena should not be dismissed as
peripheral, as is standard.
11 As in the case of linguistic names, each wh expression is taken to project a uniquely
distinguished variable of the form WH, so in eect we are assuming an open-ended set of such
meta-variables. We ignore this here.
12 This lexical specication of an annotation to a node requiring type t with a DU formula
Cat(Q) is the lexical analogue within this framework of the wh criterion: see May 1985, Rizzi
1990.
162 Wh Questions: A General Perspective
(5.46) Which director, who has directed which lms, did Sophia Loren
marry?
(5.47) Which famous lm star did the Queen, who was visiting which
hospital at the time, insist on seeing?
(5.48) Which famous dog-show in the US was visited by the Queen, who
bought how many corgis there?
Wh -in-situ forms in Iraqi Arabic do not follow this pattern, because they re-
quire a locality restriction analogous to that of re
exive anaphors. This can be
characterized in terms similar to the locality restriction formulated for pronouns
(see section 5.4.2.3). W h expressions, e.g. sheno (= `what'), can then be taken
either to annotate an unxed node, or to annotate a xed node, the latter rela-
tive to the locality constraint that the feature Q be local (the characterization,
"0 "1 (Cat(Q)) in the lexical entry below):13
sheno
IF f?T y(e)g
THEN IF fh" i(?T y(t)); ?9xT n(x)g
THEN put(F o(WH); [#]?);
go(h" i);
put(Cat(Q))
ELSE IF f"0"1 (Cat(Q))g
THEN put(F o(WH); T y (e); [#]?)
ELSE ABORT
ELSE ABORT
We have, that is, the analogue of an anaphor for the WH meta-variable { a
wh expression projecting an interpretation only in a certain most-local environ-
ment.14
It should be noted that this characterization depends on the presence of a
feature Q annotating all top nodes of wh question structures; and in making this
assumption, the feature Q is used as a more general classicatory device in Iraqi
Arabic than that for which it was previously used in characterizing languages
such as English and German. In Iraqi Arabic, given the locality restriction
associated with the wh forms themselves, it is essential that the Q feature pick
out the whole class of wh questions. In Germanic languages, to the contrary,
given the particular properties associated with clause-initial wh expressions, it
13 The argument put forward by Ouhalla (1996) that Iraqi Arabic wh -in-situ questions must
have no feature Q on the grounds that if they did, movement would be enforced, as it is in
wh -initial structures, is theory-internal and does not carry over to this framework. In virtue
of the locality eects observed here, Ouhalla argues that wh expressions are anaphors with
the feature +WH as antecedent. The analysis given here preserves the spirit of anaphoricity
which Ouhalla's analysis captures in rather dierent terms.
14 It is arguable that all meta-variables have both locally restricted forms and forms which
are not so locally restricted. See Progovac 1991 for evidence that negative polarity forms
in Serbo-Croatian may range over those which are analogous to anaphors and those which,
conversely, have an anti-locality restriction.
Expletive wh Structures 163
is those fronted wh constituents and not the whole class of wh questions which
the classicatory feature Q picks out.15
5.4 Expletive wh Structures
A further type of wh expression is the so-called expletive form of wh, occurring
in German, Hungarian and Iraqi Arabic and serving as an anticipatory indi-
cation of a wh question with a full wh expression to follow.16 This is where
the relationship between requirements on a node and the subsequent successful
annotation of that node can again be put to use { wh expletives, we propose,
project a requirement for an annotation which only a corresponding full wh
form can supply. It is this relationship between annotations and requirements
for those annotations which provides a basis for explaining the puzzles posed
by expletive wh phenomena and a structure called `partial movement' which in
some languages is characteristic of these expletive wh forms.
5.4.1 Partial Movement
Expletive wh constructions and the process of partial movement go together,
because the latter is part of a movement-based explanation of expletive wh forms
for languages such as German in which the primary form of wh expression occurs
clause-initially. The presence of an expletive wh form, which itself occurs clause-
initially, has the eect of not only inducing a lower clause-initial wh expression,
but also ensuring that that wh expression does not occur in the position it would
have done in the absence of any such expletive expression:
(5.49) Was glaubst du, wen Jakob half? [German]
What think you whom Jacob helped
`Who do you think Jacob helped?'
In (5.49) was is the expletive, wen the `partially moved' full wh form following
a verb which need not itself select for question forms. Expletive wh phenomena
also occur in languages where the wh form characteristically occurs in an in-situ
position. Iraqi Arabic, for example, has an expletive wh form, prexed on the
verb of the higher clause which licenses construal of the wh in a subordinate
nite clause, which is otherwise debarred:
(5.50) sh-'tsawwarit Mona Ali raah weyn [Iraqi Arabic]
Q thought Mona Ali went where
`Where did Mona think that Ali went?'
15 Though we do not explore the consequences of this language-particular use of syntactic
features here, it re
ects the fact that in Iraqi Arabic, the in-situ position is the unmarked
form of wh question when this position is licensed, whereas in the Germanic languages, it is
the fronted wh-initial structure which is the unmarked form of wh question.
16 They also occur in Hindi (see Mahajan 1990, Dayal 1994), though we leave Hindi on one
side. In that language the apparently expletive form is anticipatory of any arbitrary type
of question, suggesting a further application of a syntactic feature QUE which picks out the
broader class of questions, to include yes-no questions (see Dayal 1994, 1996).
164 Wh Questions: A General Perspective
The encliticized form sh- is a device which explicitly licenses the presence of a
wh expression in the next tensed clause, and does not license wh expressions
across an indenite number of tensed clause boundaries:17
(5.51) *sh-i'tiqdit Mona meno tsawwarit Ali sa'ad meno
Q believed Mona who thought Ali helped who
`Whoi did Mona believe ei thought Ali helped who?'
Unlike Iraqi Arabic, German allows partial movement with an indenite number
of expletive was expressions followed at some lower clausal boundary by the full
wh expression, which must occur initially in the clause:18
(5.52) Was glaubst du, was Hans meint, wen Jakob liebt?
What think you what Hans said whom Jacob loved
`Who do you think Hans said Jacob loved?' [German]
Partial movement is a long-distance dependency eect, since not only the ex-
pletive form but also the full wh expression is free to occur arbitrarily far from
the ultimate gap position:
(5.53) Was glaubst du, wen Hans meint, dass Jakob half?
What think you whom Hans said that Jacob helped
`Who do you think Hans said Jacob helped?'
In some dialects of German, the anticipatory sequence of was expletives must be
unbroken, with each expletive only able to license a wh expression across a single
tensed clause boundary. In others, however, speakers report the acceptability
of a single anticipatory was followed by an intervening dass:
(5.54) Was glaubst du, dass Hans meint, wen Jakob half?
What think you that Hans said whom Jacob helped
`Who do you think Hans said Jacob helped?'
What is essential is that there be no occurrence of such a was complementizer
following a full wh form, unless (as in (5.57)), there is some following second
full wh form of which the intermediate was is an anticipation:
(5.55) *Wen glaubst du, dass Hans meint, was Jakob half?
whom think you that Hans said what Jacob helped
`Who do you think Hans said Jacob helped?'
(5.56) *Wen glaubst du, was Hans meint, dass Jakob half?
Whom think you what Hans said that Jacob helped?
`Who do you think Hans said Jacob helped?'
17 The sh- form is an abbreviated form of sheno meaning `what'.
18 Examples are taken primarily from Simpson 2000 and Hohle 1996.
Expletive wh Structures 165
(5.57) Wer glaubt, was sie meint, wen Hans liebt?
Who believes what she said whom Hans loves
`Who believes whom she said Hans loves?'
In German, other wh expressions also have apparently expletive uses, though
these require successive occurrences of the particular expletive form, without
any intervening dass:
(5.58) Wer glaubst du, wer Uli half?
Who think you who Uli helped
`Who do you think helped Uli?'
(5.59) Wen meint Karl, wen Uli half?
Whom said Karl whom Uli helped
`Who does Karl think Uli helped?'
(5.60) Wie nimmt man, an wie der Prozess ausgeht?
how assumes one how the process ends
`How does one assume that the trial ends?'
Though the expletive form is unambiguously construed as a marker of some
lower wh expression, as re
ected in the case assignment to these German ex-
amples, it may on the other hand display case properties which re
ect its own
position in the lexical sequence. For example, in Hungarian, which also displays
partial movement, the expletive form is case-marked not by the embedded posi-
tion corresponding to the interpretation to be projected, as in German, but by
the predicate local to its occurrence in the string:
(5.61) Mit mondtal, hogy kinek vett Janos
WhatACC said2:sg:indef:ACC that whoDAT bought JanosNOM
szinhazjegyet? [Hungarian]
theatre-ticketACC
WhatACC did you say for whom John bought a theatre-ticket?
`For whom did you say John bought a theatre-ticket?'
Note in (5.61) the accusative marking on the expletive form mit, and not the
dative, re
ecting the resulting position of the WH variable.
5.4.1.1 Minimalist Accounts: Direct versus Indirect Dependency
Faced with this amount of variation, it is perhaps not surprising that heterogene-
ity in the analysis is taken to be required; and current Minimalist explanations
are divided over the underlying uniformity of the data (see Horvath 1997, Dayal
1996, Beck and Berman 1996 for arguments that the phenomena require more
than one form of explanation). But the real puzzle for Minimalist assumptions,
as is demonstrated in detail by Simpson (2000), is why partial movement should
exist at all. Given that the expletive ensures (e.g. through movement to a Spec
166 Wh Questions: A General Perspective
CP position) that the wh feature is checked under adjacency, there should be
no need of movement of the full wh form, and with no need of movement, no
movement should be licensed. It is thus far from clear how the obligatory nature
of partial movement in a language such as German can be licensed, let alone
forced. To make matters worse, such partial movement should be to a node
which is marked for the feature Q, but partial movement constructions char-
acteristically exist as complements to predicates which select as complements
propositions which are not questions, hence with the feature Q.19
The various Minimalist analyses of these data can be seen as attempts to
overcome this problem { none of them fully satisfactory. There are two princi-
ple variant forms. One entails analysing the anticipatory wh expression as some
form of expletive, creating a single chain from the expletive form to the wh form
(and thence to its gap). This direct dependency approach is so called because
it directly re
ects the interpretation, which corresponds to long-distance depen-
dency construals as in English (McDaniel 1989, Chomsky 1995, Brody 1994,
Beck and Berman 1996). The other, so-called indirect dependency approach
analyses the expletive forms as question words in their own right, each clause
corresponding to an independent question with local binding of some posited
empty position for the expletive wh form (Dayal 1996, Horvath 1997). The
direct dependency approach is clearly inappropriate for Hungarian, since case
assignment in these Hungarian structures must not re
ect the position of the
gap. However, the indirect dependency approach cannot be seen as a simple
apposition of two entirely independent questions, since a quantier within the
initial clause may bind a pronominal in the second (Beck and Berman 1996,
Dayal 1996, Horvath 1997):
(5.62) Was glaubt jederi , wohin eri geht? [German]
what thinks everyone where he went
`Where does everyone think he went?'
(5.63) Mit nem hisz senki, hogy milyen torteneteket
WhatACC not believe no-one that what stories
terjeszt rola a felesege [Hungarian]
spreads about him the wife-hisNOM
`What stories doesn't anyone believe that his wife spreads about him?'
Moreover, there is the problem of what should be taken as the trigger of the
two movement processes that have to be invoked { on the one hand, partial
movement to a Spec CP position whose CP is immediately dominated by a
predicate which precludes questions, and, on the other, a subsequent pied-piping
at LF, which is needed to ensure the interpretation of the whole as a single
question.20
19 This problem applies to all analyses in which the feature Q is taken to be a feature
indicating the semantic type of the clause, as in Brandner 1996.
20 Pied-piping at LF, which is highly problematic under Minimalist assumptions, is also
needed in some direct dependency analyses (see Fanselow 1997).
Expletive wh Structures 167
Even leaving aside how to dene appropriate triggers, there is the problem
that indirect dependency approaches fail to explain why no expletive forms occur
as second or subsequent members of a multiple wh construction:21
(5.64) *Wer hat was gedacht, wen wir anrufen sollen? [German]
Who has what thought whom we call up should
`Who thinks what who should we call up?'
There is also the problem that the relation between the expletive wh form and
its supposed site does not parallel wh movement. As Horvath points out, in
Hungarian a wh determiner may occur externally to its DP, indicating movement
across a DP boundary, an option which is not available to the expletive wh
forms:22
(5.65) Kinek lattad a feleseget? [Hungarian]
WhoDAT saw2sg:def:DO the wife3sg:poss:ACC
`Whose wife did you see?'
(5.66) *Mit lattal kinek a feleseget?
WhatACC saw2sg:indef:DO whoDAT the wife3sg:poss:ACC
`What did you see whose wife?'
The expletive also cannot be followed by the full wh expression embedded in
a non-nite clause, despite the fact that movement across such a boundary is
unproblematic:
(5.67) Kivel akarsz beszelni?
Who-with want2sg:indef:DO talkINF
`With whom do you want to talk?'
(5.68) *Mit akarsz kivel beszelni?
WhatACC want2sg:indef:DO who with talkINF
What do you want with whom to talk
`With whom do you want to talk?'
This restriction is quite unlike wh movement, and requires special explanation.
We should notice in passing that it echoes the Iraqi Arabic data in which the
function of the expletive is to signal the presence of a wh expression in a locally
subordinate tensed clause. And all of these indirect dependency approaches are
inappropriate for the German non-was forms of expletive (5.58){(5.60), which
have to be set aside as involving a separate pronominal form of wh expression
21 Simpson (2000) suggests a case explanation, but this leads to the consequent need in
explaining Hungarian sentences analogous to the German (5.54) (in which there is an inter-
vening non-expletive complementizer) to postulate a PF-driven movement of an expletive form
in Hungarian of which there may be no phonetic realization.
22 Hungarian marks deniteness and indeniteness constraints on the object as a suÆx on
the verb, here indicated as e.g. `def.DO'.
168 Wh Questions: A General Perspective
despite the parallelism observed between these and the more regular was form
of expletive (see McDaniel 1989, Reis 1996).23
It is in the light of these and other problems that some authors grant the
possibility of setting up quite dierent types of analysis for expletive forms, with
German said to display a direct dependency type of structure, Hungarian an in-
direct dependency form of structure (see Beck and Berman 1996, Horvath 1997,
Dayal 1996). The current overall position in Minimalism debates about partial
movement thus presents a lack of consensus as to how to resolve the problems
these structures pose, with no very strong justication for the movement of the
intermediate form and no apparent means of characterizing the various dierent
case realizations allowed across dierent languages in a uniform way.
5.4.1.2 HPSG { A Feature-based Account of wh Variation
The
exibility of feature specications provided by the HPSG formalism allows
the wh phenomena to be characterized relatively unproblematically (Johnson
and Lappin 1999), though the formalism's lack of intrinsic dynamism makes it
not well suited to explain why the data should pattern as they do. The diversity
in the data of long-distance dependency, wh -in-situ and expletive wh forms is
simply carried over to the analysis itself. Three separate binding operations
are dened for the dierent structures: dislocated wh expressions, in-situ wh
expressions, and expletive wh expressions. Regular wh structures are character-
ized by feature percolation of SLASH features to be discharged by the abstract
Q complementizer in the presence of the initial wh expression; wh -in-situ struc-
tures are characterized as projecting a distinct feature QUE, also discharged by
the abstract Q complementizer. The Q complementizer itself involves a three-
way disjunctive specication to embrace rigidly wh -in-situ languages, mixed
languages that display both wh -in-situ and `moved' forms, and languages which
allow all wh forms to cluster at the front of the clause. The expletive wh is anal-
ysed as itself a complementizer that discharges a newly dened feature LQUE, a
feature that is introduced by a second abstract complementizer in the presence
of the full wh expression and subsequently passed up the tree. Each form of
feature percolation is subject to discrete forms of island, with additional locality
23 A pronominal copy cannot provide any basis for predicting the sharp fall-o in accept-
ability if the number of occurrences exceeds two, an asymmetry with was which is directly
contra-indicated by a pronominal account of these additional non-was expletive forms.
(i) ?*Wer glaubst du, wer Hans meint, wer Uli half?
`Who do you believe Hans said helped Uli?'
(ii) Was glaubst du, was Hans meint, was Jakob sagt, wen Uli gesehen hat?
`Who do you think Hans said Jacob said Uli has seen?'
(i) is markedly worse than (ii), notwithstanding a form of wh, which, on a pronominal copy
analysis, should make the sequence as easy to process as, say, (iii):
(iii) Jeder student weiss dass er seine Hausaufgaben zu seinem Lehrer schicken hatte sollen
wenn er sein Ergebnis rechtzeitig erhalten wollte.
`Each student knows that he should have sent his assignments to his teacher if he
wanted to know his results in time.'
Expletive wh Structures 169
features specic to each construction to induce whatever restrictions are neces-
sary (island restrictedness in the rst, not so restricted in the second, possibly
locally restricted to a single clausal boundary in the third). The particular as-
sociation of type of restriction to each feature is coincidental. A SLASH feature
insensitive to islands and a wh -in-situ expression sensitive to such islands could
equally be dened. Where the expletive form itself displays case, as in Hungar-
ian, a discrete analysis would be required in order to re
ect the case feature,
a prediction not made by the LQUE analysis. So the account succeeds, but it
does little more than mirror the divergence of the data, with no explanation of
the data in terms of independently motivated general principles.24
5.4.2 Partial Movement as a Re
ex of a Requirement
In contrast to these exclusively conguration-based approaches, the dynamics of
the present framework might have been devised with the expletive wh phenom-
ena in mind; for the shift from an imposed requirement to a resulting annotation
{ in combination with the concept of nodes in a tree being describable as unxed
{ is, we suggest, what is needed to capture the underlying uniformity in these
data.
We start from the simplest case { a regular wh expression annotating directly
a xed position with Fo(WH). If we were to dene a specic lexical item to
ensure the subsequent presence in the string of such an expression, and in this
sense anticipate it, this item would have to project a requirement:
? # (F o(WH ))
To ensure that such an expression would annotate a node properly dominated
by the node under construction, the requirement would have to take the form:
?h#i # (F o(WH ))
This gives us the means of characterizing an expletive for a wh-in-situ form, as
we shall see for Iraqi Arabic.
Supposing, however, we take the set of annotations projected by a clause-
initial wh expression. In wh-initial questions in English and in German such
expressions simultaneously annotate two nodes: the node requiring type t which
it annotates with the feature Q, and an unxed node, which it annotates with
Fo (WH). Given that any DU formula may be imposed on a node as a require-
ment, the eect of this set of actions can also be imposed as a requirement.
We will construct this requirement in stages. Seen from the perspective of some
node requiring ?T y(t), an unfixed node annotated with F o() can be described
as h#i(F o()) (note the use of the internal operator, as opposed to the external
operator used in the requirement for the in-situ expletive). However, wh-initial
expressions in questions also add the annotation Q to the node from which they
24 So far as we are aware, there is no categorial grammar account of partial movement
phenomena.
170 Wh Questions: A General Perspective
dangle as unxed. At this node, the eect of this pair of actions can be described
as:
Cat(Q) ^ h# i(F o(WH ))
As in the simpler wh-in-situ case, we have to ensure that this requirement can
be fullled only by wh expressions annotating a discrete, dominated node.
Thus, we need a requirement of the form:
?h#1i # (Cat(Q) ^ h#i(F o(WH )))
This gives us the requirement needed to characterize an expletive which is antici-
patory of a clause-initial wh expression, as is needed for German and Hungarian.
This requirement is satised only in the presence of two nodes (apart from
the one it decorates): the unxed node and a node on the branch leading to
that unxed node. In principle, either of the two nodes to be introduced may
be subject to further restrictions { by providing a more precise description of
their relative position in the tree (case marking) { or in terms of their relative
distances. This gives us directly the span of variation we nd in expletive wh
forms: German optionally case-marks the unxed node, indicating its relative
position in the resulting tree, with variation over how local the intermediate
node needs to be to the node decorated by the requirement. Hungarian case-
marks the intermediate node.
5.4.2.1 German wh Expletives
German, like English, denes its full wh expression, if initial, as not only
annotating an unxed node with Fo(WH), but also annotating the dominating
node requiring T y(t) with the additional formula Cat(Q). Accordingly,
the expletive was form is dened to impose the composite requirement
Cat(Q) ^ h# i(F o(W h) ^ T y(e)) on some dominated node. However, in or-
der to dene the update provided by the expletive form, we need to bear in
mind appropriate word order restrictions, as it is the lexicon which has to de-
termine obligatory aspects of word order. So before turning to this task, we
provide a sketch of German word order and the general assumptions we shall
make in addressing the problems raised.
5.4.2.1.1 German word order German presents notorious problems of
word order. Primary amongst these is the verb-second phenomenon: in root
clauses German allows almost any constituent to be initial, but following this
constituent, the verb must occur on its own, leaving the subject and any other
noun phrases to follow (in the so-called Mittelfeld), in which they occur in the
canonical order subject/indirect object/direct object:25
25 There are minor exceptions to this verb-second restriction: e.g. topic constructions, for
which an analysis of the rst and second constituents as the root node of a LINK ed structure
and an unxed node respectively seems promising:
Expletive wh Structures 171
(5.69) Den Marie liebt der Jakob. [German]
MarieACC loves JacobNOM
`Jacob loves Marie.'
(5.70) Der Jakob liebt den Marie.
JacobNOM loves MarieACC
`Jacob loves Marie.'
To add to the complications, if such a verb form is compound, as in the perfect,
then it is the auxiliary which occurs in the second position in root clauses, and
in such a case the verb must occur at the end, following the sequence of noun
phrases in the Mittelfeld:
(5.71) Heute hat der Jakob den Marie gesehen.
Today has JacobNOM MarieACC seen
`Today Jacob has seen Marie.'
(5.72) Den Marie hat der Jakob heute gesehen.
MarieACC has JacobNOM today seen
`Marie, Jacob has seen today.'
(5.73) *Der Jakob hat gesehen den Marie heute.
JacobNOM has seen MarieACC today
`Jacob has seen Marie today.'
(5.74) *Heute, der Jakob hat gesehen den Marie.
Today JacobNOM has seen MarieACC
`Today, Jacob has seen Marie.'
This nal position in a clausal sequence is the required position for verbs in
subordinate clauses, unless the verb takes a complement clause, in which case
in both root and subordinate clauses the complement clause must follow the
verb:26
(5.75) weil der Jakob den Marie liebt
because JacobNOM MarieACC loves
`because Jacob loves Marie'
(5.76) Hans sagt, dass er glaubt, dass der Jakob den Marie liebt.
Hans said that he thinks that JacobNOM MarieACC loves
`Hans said that he thinks that Jacob loves Marie.'
(i) Den Jakob, den liebt die Marie.
JacobDAT , thatDAT loves MarieNOM
`Jacob, Marie loves.'
Dialects of German vary as to whether case is obligatory with names. Here we give examples
with case marking on all names, though case marking of names, in particular when nominative,
is often avoided.
26 Again with minor exceptions, in which a subordinate clause may be taken to mimic the
root clause ordering.
172 Wh Questions: A General Perspective
In wh questions, the initial position in root clauses is always taken by the wh
expression (which is, as in English, its canonical position), and in subordinate
clauses, the wh expression must, as in English, occur rst in the sequence of
noun phrases:
(5.80) *Was wer glaubt, dass der Jakob den Marie liebt?
What who thinks that JacobNOM MarieACC loves
`Who thinks that Jacob loves Marie?'
fT n(b); ?T y(t); }g
fT n(a); ?T y(t)g
feature Q, and from then on, the construction mirrors the one for wh-initial
structures. The nal result is that two strings diering only with respect to
the presence of an expletive plus an intermediate clause-initial full form versus
full form in initial position will yield the same tree structures. Both (5.81) and
(5.82),
(5.81) Wer glaubst du, dass Uli half?
Who think you that Uli helped
`Who do you think helped Uli?'
(5.82) Was glaubst du, wer Uli half?
What think you who Uli helped
176 Wh Questions: A General Perspective
`Who do you think helped Uli?'
will yield the same annotation of the root node:
Cat(Q); F o(Glaub(Half(Uli)(WH))(du)); T y(t)
Either one of the two nodes required by the expletive might be additionally
restricted. For example, to the requirement of an unxed node annotated with
Fo(WH), one might add a case specication constraining its eventual position
in the tree. This yields the additional type of German expletive, the lexical
variants of wer, wen, etc. (repeated here from section 5.4.1):32
These cases, licensed in some dialects of German, also illustrate the second basis
for variation outlined earlier, that of constraining the tree relation between the
node decorated by the expletive and the one annotated by Fo(WH). In these
forms, the relation between the node decorated by the expletive requirement
and the dominated node with the feature Q is restricted by the fact that there
may be only one intervening type t node. This more restricted relation, together
with some case specication such as subject, can be expressed as:
?h#1i #1#0 (Cat(Q) ^ 9xT en(x) ^ h#i(F o(WH) ^ "0 (T y(t))))
The case-marked expletives constrain their tree positions. We give here the
lexical characterization of expletive wer:33
32 In movement analyses of expletive wh phenomena in German, these are referred to in
the literature as an entirely separate `copy-wh strategy', on that analysis the broad range of
properties (see Hohle 1996, Reis 1996) shared by the two processes being coincidental.
33 Here we are systematically ignoring person considerations, though their specication
might be incorporated by a restricted indexing system. See the lexical account of the denite
article and the encoding of its restrictor properties suggested in chapter 7.
Expletive wh Structures 177
werexpl
IF f?T y(t)g
THEN IF f# 9xF o(x)g
THEN ABORT
ELSE IF f" >g
THEN put(Cat(Q),
?h#1i #1#0 (Cat(Q) ^ 9xT en(x)^
h# i(F o(WH) ^ h"0 i(T y(t))));
make(h#0 i); go(h#0 i)
ELSE put(Cat(Q),
?h#1i #1#0 (Cat(Q) ^ 9xT en(x)^
h# i(F o(WH) ^ h"0 i(T y(t)))));
make(h#1 i); go(h#1 i); put(?T y (t ! (e ! t)))
ELSE ABORT
Notice how this form, like the expletive forms of wen and wem, display system-
atic variation from the lexically neutral was on two counts: case and locality,
and also in fact person marking, which we have ignored here.
By way of comparison we give in gure 5.3 schematic output descriptions
for (5.81), (5.82) and (5.58), each of which gives rise to the same formula. Note
that the rst characterization involves two nodes, but the two following only
one; the third imposes a case requirement at some unxed node annotated by
Fo(WH), and a locality requirement on the intermediate node.
All these German wh expletives share overall properties. On the one hand,
the Q feature which they impose as a requirement on a dominated node will not
clash with the semantic property of being the object of a propositional attitude
(which is required of complements to predicates such as glaub ), since Q is indica-
tive of the structural property of a formula containing the meta-variable WH, not
of any general property of questionhood. The signicance of this is that, on the
one hand, there is no con
ict with whatever semantic restrictions are imposed
by intervening verbs such as glaub, which may not select for questions.34 On the
other hand, there is no problem in having a tree structure of which several nodes
are annotated with the feature Q, say re
ecting the formulae Liebt(WH)(Jakob)
at one node, the formula Meint(Liebt(WH)(Jakob))(Hans) at some higher node,
and the formula Glaubt(Meint(Liebt(WH)(Jakob))(Hans))(du) at some node
which dominates both, all in re
ection of a sequence of expletives followed by a
single wh expression, each of which annotates a node with the feature Q:
(5.52) Was glaubst du, was Hans meint, wen Jakob liebt?
34 Unlike wonder in English, glaub imposes no requirement other than ?T y(t) on its induced
complement argument node, though nothing precludes its acquiring a feature such as Q.
178 Wh Questions: A General Perspective
fCat(Q); F o(Glaubt(Meint(Liebt(WH)(Jakob))(Hans))(du)g
...fCat(Q); F o(Meint(Liebt(WH)(Jakob))(Hans))g
...fCat(Q); F o(Liebt(WH)(Jakob))g
Crossover Phenomena
In exploring the dierent types of relative clauses in chapter 4, we had a rst
indication that anaphora resolution can play a central role in a structural con-
struction process, while retaining its essentially unrestricted, hence pragmatic,
nature. In this chapter, with the characterization of wh questions now also es-
tablished, we turn to so-called crossover phenomena, which have been treated
as constituting a heterogeneous set needing restrictions which are independent
both of anaphora construal and of long-distance dependency. Even a subset of
these phenomena has been described by Postal (1993) as a mystery, split into
an increasing number of discrete sub-phenomena. Together they constitute a
phenomenon in which anaphora and wh construal evidently interact; and the
challenge will be to provide a unifying analysis while nevertheless respecting the
diversity of possible interpretations.
6.1 Crossover { The Problem
In the 1970s, as a movement account of wh expressions was under development,
it was noted by Postal (1972) that a pronominal cannot be identied with a wh
expression (in (6.1)-(6.3) the relative pronouns) if it occurs between the position
in which that wh expression occurs and the position at which the wh expression
is to make its contribution (in the movement metaphor, the gap):1
(6.1) *Johni, whoi Sue thinks hei worries ei is sick, is on half pay.
(6.2) *Every man whoi Sue thinks hei worries ei is sick, is on half pay.
This phenomenon extends equally to wh questions:
(6.3) *Whoi does Sue think hei worries ei is sick?
1 In this chapter and elsewhere, we indicate this position by co-indexing of ei with the
appropriate expression only when either reporting on movement analyses or when it is essential
for clarity of the exegesis.
190
Crossover { The Problem 191
The rst problem posed by these data is that they seem to show that wh ex-
pressions cannot be analysed as regular quantier-variable binding expressions,
because, if they were, they would be expected to bind any or all variables within
their c-command domain. The solution, uncontroversially, is taken to be a dis-
crete denition of a syntactic operator. The second problem the crossover data
pose is that a subset of the data appears to be sensitive not merely to the rela-
tion between the wh element, the pronominal and the position of the so-called
gap, but to the larger context in which that structure is contained.2 Despite the
apparently identical congurational relationship between wh, pronominal and
gap in relatives and questions, the crossover eect appears not to be constant
across the two structures. In relatives, the restriction debarring an intervening
pronominal appears to hold with much reduced strength, if at all, if either the
wh or the pronominal expression is nested within a determiner:3
(6.4) The studenti whoi Sue thinks hisi mother worries ei is sick is on half
pay.
(6.5) The studenti whosei mother Sue knows hei worries ei is sick is on his
way to the hospital again.
(6.6) The book whichi I had to tell itsi author ei wasn't selling well was
taken o the market.
(6.7) Everyone whoi hisi mother had helped ei did well in the exam.
(6.8) The nurse whosei patientsj shei admitted ej were unhappy, was
transferred to another ward.
(6.9) Any student whosei abstractj the committee failed to tell themi ej
had been accepted, had assumed itj had been rejected.
This asymmetry between the pairing of a pronominal and a wh expression both
to be construed as arguments and the pairing of pronominal and wh when at
least one of them is not an argument does not apparently extend to questions,
where the unacceptability of an intervening pronominal appears to be sustained
for all sequences of wh ... pronoun ... gap:4
(6.10) ?*Whoi does Sue think hisi mother worries ei is sick?
2 The crossover restriction is said also to underlie the unacceptability of (i), a constraint
holding of movement at LF, but following Williams (1994), we take this to be a linearity
restriction, and address only the interaction of pronominal anaphora and wh expressions:
(i) *Hisi mother annoyed every boyi .
3 It is reported in the literature that in relatives, it is only the non-restrictive form in which
weak crossover eects are suppressed, but in view of the fact that we have not met a single
speaker who does not accept all of (6.4){(6.9), we shall take the primary data to be those in
which weak crossover and extended strong crossover eects are suppressed whenever either
the wh expression or the pronominal is a sub-part of a determiner.
4 We shall see in due course that the data are not this clearcut. For some speakers, questions
with which are also acceptable under the indicated reading:
(i) ?Which childi is hisi mother looking for?
192 Crossover Phenomena
lack of any island eects with wh resumptive pairings was rst observed). How-
ever, this promise of a solution by simply providing two modes of analysis has
not provided the easy, albeit weak, solution for the crossover mystery that Sar
promised.5 To get the analysis to work, Sar has to posit a principle whereby a
syntactically derived gap (= a derivational variable) is converted to a resumptive
pronoun whenever it is not bound to a true quantier, and the relative pronoun
in non-restrictive relatives is transformed from being an operator at LF to a
referential expression at LF0 (this presumably despite current assumptions that
LF is the level at which interface properties with the semantics are dened).
As Aoun and Choueiri (1997) and Aoun and Benmamoun (1998) have pointed
out, the pairing of a dislocated expression with a resumptive pronoun is in any
case not clear indication of a non-movement analysis, for there is interaction in
all forms of Arabic between left-dislocation eects associated with a gap and
left-dislocation eects with a pronoun if both occur within the same string. If a
dislocated NP and its paired resumptive pronoun are contained within the con-
guration intervening between a wh expression and its gap, then the dislocated
NP and its paired resumptive pronoun become subject to the very same restric-
tions as those of the wh expression and its gap, viz. sensitivity to islands. Thus,
despite the acceptability of (6.23) and (6.24), (6.25) and (6.26) are reported to
be quite unacceptable:
(6.23) Drna
y l-masra iyye yalli t?arrafna ?ala l-muXriz
y p p yalli
saw1st:P lur the play that met1st:P lur on the director that
?aXraz-a
p [Lebanese Arabic]
directed3:Sing:Masc it
`We saw the play that we met the director that directed it.'
(6.24) t?arrafna ?ala l-muXriz yalli Laila seet l-masrahiyye
p p
(6.33) Johni, whosei auntj hei knows Mary likes ej , is insisting on ignoring
her.
(6.34) Johni, whosei auntj ej knows Mary likes himi , is insisting on ignoring
her.
(6.35) Johni, whosei auntj hisi sister knows Mary likes ej , is insisting on
ignoring her.
The reason for this emerges from considering the successive steps of interpreta-
tion.
We take, rst, the interpretation of (6.28), paying close attention to the
construction steps. In (6.28), the relative clause is initiated by the imposition
of a requirement that within the resulting representation there must be some
node annotated with the formula F o(John) (through LINK Introduction ). The
immediately subsequent processing of the relative pronoun annotates the un-
xed node bearing this requirement, so that all that is left is to establish a xed
tree position. We have, that is, the partial tree description which we might
display as in gure 6.1. The next word to be interpreted in the sequence is he,
which, in virtue of its case specication, is established as annotating a subject
node (constructed through a step of Introduction and Prediction ). The meta-
variable associated with this pronominal requires a substituend, and a candidate
antecedent formula is F o(John) annotating the node of the previously estab-
lished incomplete tree described as fh"0iT n(a); F o(John); T y(e)g. However,
this annotation cannot be chosen. Interaction between an unxed node anno-
tated by John and the locality constraint on Substitution associated with the
198 Crossover Phenomena
fT n(a); ?T y(t)g
John
fT n(a); ?T y(t)g
fh"0 iT n(a);
F o(John); T y (e)g fh"1 iT n(a); ?T y(e ! t)g
John
fh"0 ihL 1 i
h"0 iT n(a); fh"1 ihL 1 i
fh" ihL 1 ih"0 iT n(a); ?T y(e)g F o(John); T y (e)g h"0 iT n(a);
?T y(e ! t); }g
he
fT y(e ! cn);
fT n(m); T y(e); F o(Aunt)g
F o(U)g
aunt
fhL 1 iT n(m);
F o(P OSS (John;U));
T y (t)g
fF o(John);
T y (e)g fF o(P oss(U)(John));
T y (e)g
whose
the man that I think that about him you said that Sara
katva sir
wrote poem
`the man that I think that about him you said that Sara wrote a poem'
(6.50) ha-?is
p se ?alav
p ?ani xosev se
p ?amarta se
p Sara
the man that about him I think that you said that Sara
katva sir
wrote poem
`the man that about him I think that you said that Sara wrote a poem'
(6.51) ha-?is
p se ?ani xosev se
p ?amarta se
p ?alav
p Sara
the man that I think that you said that about him Sara
katva sir
wrote poem
`the man that I think that you said that about him Sara wrote a poem'
The explanation turns on the fact that there are two aspects of underspeci-
cation in the projection of LINKed structures. On the one hand, there is the
requirement for the construction of a copy of the formula from the head node.
On the other hand, there is the requirement for some tree node to host this re-
quired annotation. Though these requirements may be met by the projection of
a xed node duly annotated directly (by suitable construal of a pronoun taken
to annotate a xed tree node), it is also possible for these two requirements to be
met separately. In particular, the construction of a copy of the required formula
can be met by first building an unxed node and decorating it. The second
step would then be nding a xed position for this node. This two-stage sat-
isfaction of a modal requirement initially imposed by LINK Introduction takes
place when the pronoun is taken to annotate a position which is itself unxed.11
We have the transition from
fhL 1 iT n(a); ?T y(t); ?hDi(F o(); T y(e))g
11 As Demirdache (1991) points out, reporting Borer (1984), topic constructions are possible,
albeit marked, within a relative clause in Hebrew (?et is a free standing morpheme which
p
to
fhL 1 iT n(a); ?T y(t); ?hDi(F o(); T y(e))g; fh"ihL 1 iT n(a); ?F o(); T y(e)g
as the rst step,12 and the transition to
fhL 1 iT n(a); ?T y(t); ?hDi(F o(); T y(e))g; f" hL 1 iT n(a); ?F o(); T y(e)g
as the second step. Note that the external operator " replaces the internal
operator h"i, describing the replacement of an unxed node by some xed
node, hence a monotonic form of update. We thus derive the dierence between
the English case (6.52) and the Hebrew (6.49){(6.51):
(6.52) *Johni, who Sue said hei knew Bill disliked ei, refused to stand for
president.
In (6.52) the interpretation for the pronoun is projected as part of the tree
structure at a xed node of type e, precluding any subsequent gap. In (6.49){
(6.51) the pronoun projects a formula onto a node, but the position of that
node in the description is not xed. This satises the modal requirement, but a
further step is required to identify a xed position for this node within the tree.
6.2.1.3 Apparent Crossover Eects in Hebrew
A second property of Hebrew is the marked weak crossover eects observed in
restrictive relatives, reportedly as strong as strong crossover eects in English
(see Demirdache 1991):
(6.53) *ha-?is sei ?im-oi
p p ?ohevet ei
p
which se annotates an unxed node, the locality restriction constraining the
interpretation of the pronominal will preclude the appropriate substitution, so
the only possible way to assign the requisite interpretation to the pronominal is
by Merge. But with the unxed node having merged with the node annotated
by the pronominal, the projection of the tree will thereafter remain incomplete
in the presence of a subsequent gap. Either way, that is, the presence of the
suÆx o within the construct state structure will have the eect of establishing
a xed position in the tree for the required copy { hence the reported eects
that the ungrammaticality is of the same order as strong crossover in English
(and not a weaker weak crossover violation). Relative to our model, (6.56) is
well-formed, as the analysis would lead us to expect:
13 The form of construct state nouns is discrete from their independent forms, is phono-
logically incomplete, with reduced stress and consequent shifts in both vowel and consonant
forms, and cannot occur independently.
14 As this analysis re
ects, such construct state forms lack a morphological determiner for
the head noun, can be recursively complex, and allow a determiner to intervene between the
head noun and its modifying noun. They also allow a clitic-doubling form of construction
with a following PP which apparently duplicates the suÆxed clitic:
(i) beit-o sel ha-more
house his of the teacher
`the teacher's house'
This structure we could dene as involving a LINK relation, the structure projected from the
PP with sel LINKed to the nominal variable of the preceding noun{clitic sequence happening
to project the very same P OSS relation, and therefore imposing identity of interpretation on
the pronominal clitic and the following full noun phrase.
208 Crossover Phenomena
(6.56) ha-?isi
p se ?im-oi
p ?ohevet ?otoi
p p
15 We thank Ad Neelemann for this observation and the discussion to which it led.
Crossover { The Dynamic Account 209
(6.60) de man van wie Sue verwacht dat Tom (hem)
the man of whom Sue expects that Tom (him)
opzettelijk negeert
deliberately ignored
`the man whom Sue expects Tom deliberately ignored'
(6.61) de man van wie Sue verwacht dat hij ziek is
the man of whom Sue expects that he is sick
`the man whom Sue expects is sick'
However, the proposed analysis of crossover is undisturbed by such examples.
We assume that in Dutch all Merge operations are of the more restricted variant,
Gap Merge, ensuring that any left-dislocated constituent cannot be identied
with a lexical pronominal. In strong crossover environments, any pronoun in-
tervening between the projection of the unxed node and the supposed point
at which Gap Merge should apply is precluded by locality restrictions from
identifying with the annotation of the head. With the relative pronoun die,
this is the entire story. With the van wie forms, on the other hand, a discrete
LINK Introduction can be projected by van wie, imposing, as in Arabic, a re-
quirement on the top node of the structure linked to the nominal variable for
a requirement of a copy, which the pronoun provides. In the latter case, the
pronoun is given a value by Substitution. Thus, though there may be individual
expressions that allow resumptive pronouns, Dutch has no general process of
pronominal construal through a Merge operation. The general form of expla-
nation for crossover phenomena is, however, unaected: it is merely simpler,
for the locality restriction associated with Substitution suÆces to preclude a
pronoun between the antecedent16 and the point in the interpretation process at
which Gap Merge might apply.
6.2.1.5 Constraining the Liberality of the Computational System?
A major question for the proposed analysis is its liberality. By allowing Merge
to be a means of assigning an interpretation to a pronoun, all head-initial rela-
tive structures will in principle be expected to license the occurrence of lexical
pronouns interpreted resumptively in positions where a gap may occur. Yet
we have already seen that this cannot be the whole explanation. In particular,
there appears to be a very general avoidance of pronouns being resumptively
construed too close to the start of the clausal sequence:17
(6.62) ?*The man who he likes my mother has come to visit us again.
(6.63) ?*The man who my mother dotes on him has come to visit us again.
16 We have not analysed English in terms of Gap Merge because, as we shall see immediately,
the acceptability of using pronouns resumptively in English is determined by an array of
system-external factors.
17 The `?',`??' or `?*' in the examples that follow are merely crude indicators of relative
judgements of (un)acceptability.
210 Crossover Phenomena
node, for example through Gap Merge.18 However, it is hard to see how the
required restriction in Arabic could be expressed. In Arabic relatives, there is
no projection of any unxed tree node, and the oending lexical pronoun is in
competition with a meta-variable projected by the verb, both of which must
make provision for a value to be assigned from an external antecedent (their in-
dexical construal). The move to debar these data as ill-formed in either English
or Arabic would, furthermore, create the concomitant diÆculty of explaining
why such strings may be, and often are, unproblematically assigned an inter-
pretation. The alternative is to dene a system which is liberal with respect to
the data in allowing as well-formed some strings which speakers may express
varying degrees of hesitation over accepting. We opt for this route as our general
strategy. In a system with a monotonic concept of tree growth, with no possibil-
ity of retraction, it is harder to explain how some string interpretation pairings
judged as acceptable are precluded than it is to explain how some unaccept-
able pairing of string and interpretation might nevertheless be included. In the
face of data where acceptability varies according to non-structural factors, we
assume that the constraints determining acceptability are a consequence of rel-
evance or other processing considerations, retaining the characterization given
in chapter 3, that it is the triple of computational, lexical and pragmatic ac-
tions which determines available pairings of strings of the language and assigned
logical forms.
We do not oer here a detailed pragmatic account of focus and contrastive
stress with which to substantiate this methodological decision.19 If a relevance-
theoretic perspective of this sort is adopted, with its commitment to explaining
interpretation in terms of balancing inferential eects against cognitive cost,20
the exceptional resumptive use of pronouns might be seen as arising from the
speaker setting aside the consideration of cost to the hearer of parsing an unnec-
essary pronoun, in the light of wanting to convey particular inferential eects
which can only be ensured through the use of a (contrastively stressed) pro-
18 In order that the application of Gap Merge in English should not be vacuous, the more
general form of Merge would have to be precluded from applying in the same context. This
would leave the admittedly borderline acceptability of (6.62){(6.69) and (6.72){(6.77) quite
unexplained unless a discrete lexical specication was posited for the stressed form of each
pronoun, such forms being analysed as expletives that fail to provide any meta-variable to sat-
isfy the requirement ?T y(e) (thereby enabling Gap Merge to apply), but nevertheless project
a focus feature relative to which focus eects could be dened. Not only would such a char-
acterization be at best partial, excluding the unstressed resumptive use of pronouns, but it
would also be predicted to be a fully acceptable form of wh question, at least if the widely held
assumptions that the left-dislocated position is a focus position, and that fronted wh expres-
sions in wh questions are associated with focus, are retained. For discussion, see Rochemont
and Culicover 1990, Chomsky 1971, Erteschek-Shir 1997. Yet it is these structures for which
resumptive construal of pronouns is problematic.
19 See Breheny 1996 for a relevance-theoretic characterization of focus giving a representa-
tionalist alternative to Rooth's model-theoretic projection of accompanying context (Rooth
1992, 1996), in terms of sets of premises made available for inference.
20 Recall the characterization of the presumption of optimal relevance as an assumption that
the intended array of inferential eects is achievable with minimal cognitive eort necessary
to retrieve that interpretation (Sperber and Wilson 1995).
Crossover { The Dynamic Account 213
noun.21 Without any such inferential eects, such use of pronouns is avoided,
because unnecessary. Adopting the methodological preferences of a more, rather
than less, liberal system, with nothing more than a promissory note of some
forthcoming pragmatic explanation,22 we retain a formal system which licenses
resumptive use of pronouns in relative clauses and, accordingly, an account of
crossover phenomena which has this construal of pronouns as a consequence.
6.2.2 Crossover Phenomena in Questions
We have so far left entirely on one side the apparent asymmetry between
crossover phenomena in relatives and in questions. It might seem that the
decision to allow pronouns in English in all positions in which a successful ap-
plication of Merge can take place is too hasty, given that pronouns cannot gen-
erally be used resumptively in wh questions (6.82) and (6.83), and many people
provide judgements of weak crossover eects that his cannot be identied with
who in (6.84) and (6.85):23
(6.82) ??Whoi did Sue insist that we invite himi to the party?
(6.83) *Whoi did Mary like himi ?
(6.84) ??Whoi did hisi book irritate?
(6.85) ??Whoi did hisi mother wish wasn't sick?
This reported unacceptability of weak crossover construal of pronouns in ques-
tions, and the accompanying unacceptability of pronouns resumptively con-
strued, is, as in relatives, not explained by the account to hand. On the basis of
the analysis so far, we would expect that (6.84) and (6.85) allow an interpreta-
tion in which the pronoun is identied by substitution of WH for U in F o(U)
projected by his, since the possessive in English projects a LINKed structure,
and this substitution would not undercut the possibility of some subsequent
application of Merge. Furthermore, we would anticipate, just as in relatives,
that a pronoun might under certain circumstances serve to identify the position
of the initially unxed node in a wh question, thereby yielding a resumptive
construal. Yet it is very generally agreed that wh questions are obligatorily
associated with a gap, an association which is observed even in those languages
21 The question of estimating cognitive cost is not unproblematic, since there is a tension as
to what it is that contributes most to cognitive cost { the parsing of a pronoun, uncertainty
as to where Merge should take place, or delay in establishing some application of Merge.
22 This characterization of the underlying pressures upon choice of form of expression as
outside the remit of the computational system is in contrast to the Minimalist program,
where analogous restrictions in the form of Minimal Move are dened to be central to the
computational system (see Johnson and Lappin 1997, 1999, for an attack on this attempt
to assign system-internal status to any such restriction). In this respect, it is also unlike the
recent moves within LFG to adopt an optimality-theoretic perspective to characterize dierent
`strengths' in pronoun types (see Bresnan 1999).
23 As before, the `??' and '*' are indicators of acceptability judgements, rather than indica-
tors of ungrammaticality.
214 Crossover Phenomena
where the resumptive pronoun strategy is the only strategy available in rela-
tives.24 We could take this as evidence that wh questions are dened along
lines suggested for Dutch relativizer die as simply being gap-seeking. There is
reason, however, to think that the apparent display of asymmetry between rela-
tive clauses and questions with respect to the licensing of a resumptive construal
of pronouns is not a lexically dened restriction, but is rather due to the inter-
action of computational and pragmatic considerations jointly delimiting the set
of sentence-interpretation pairs licensed by the computational/lexical actions.
In eect, we shall be arguing that the phenomenon dubbed `weak' crossover
is a phenomenon where the preclusion of certain interpretations of a pronoun
is a consequence of pragmatic considerations alone, and is not an exclusively
structural phenomenon.
We take rst the strong crossover data, only to set them aside as exactly
parallel to their counterpart in relatives.
(6.86) *Whoi did Sue believe hei realized ei had given the wrong answer?
Any attempt to provide the value WH for the meta-variable U projected by the
pronoun by Substitution falls foul of the imposed locality constraint, and any
application of Merge to the node annotated by F o(U) and the unxed node
leads irrevocably to an incomplete structure. We leave these data without more
ado.
With the weak crossover data, we face the question of the degree to which a
wh expression can be used to provide an antecedent for a following pronominal
in wh questions. In this connection, the rst point to observe is that whatever
the supposed asymmetry between (6.87) and (6.88),
(6.87) ?Whoi did hisi mother defend?
(6.88) The man whoi hisi mother defended won his case.
the tendency to avoid using a fronted wh expression as antecedent for a subse-
quent pronominal cannot be put down to any lack of content for a WH meta-
variable. A wh expression in such a position can be successfully used to provide
an antecedent for construal of a pronoun within some following inserted paren-
thetical structure { and notice that once having been so used, the interpretation
of the pronominal within the following weak crossover environment becomes
fully acceptable:
(6.89) Whoi , if hei could be persuaded to stay, would hisi supervisor agree to
support?
24 Some languages which have a distinction analogous to the what/which distinction of
English are reported to dierentiate between the what/who types of form and the which
type of form, with the former requiring a gap, the latter a pronoun resumptively construed:
see Dobrovie-Sorin 1990 for discussion of Romanian. It is notable in this connection that
the Romanian wh interrogatives that require a clitic pronoun also require a particle pe, a
particle which appears to project a LINK relation between the left-dislocated expression and
the remainder, being obligatory also in relative clause sequences where a clitic pronoun is also
required.
Crossover { The Dynamic Account 215
(6.90) Whoi , no matter how long hei cries, will even hisi mother ignore?
(6.91) Whati , given that iti had been discussed at the previous AGM, did
itsi proponents decide not to put on this year's agenda?
(6.92) Which horsei, given that iti was badly injured in the recent National,
did itsi owner reluctantly decide not to race again ?
(6.93) Whosei dissertationj , if theyi complete itj in time, will you
recommend for the prize?
According to the representational account of pronominal construal advocated
in the present framework, this is as it should be. Any accessible representation
of the right logical type should be suÆcient as a substituend for the U meta-
variable projected by a pronoun; and the semantic weakness of a WH meta-
variable is not grounds for its unsuitability as antecedent for a pronoun.
As these examples also show, the availability or non-availability of a left-
peripheral wh expression as antecedent for a pronoun within a following subject
position does not seem to be structural either, but is rather due to a medley
of pragmatic factors. First, wh questions in which the pronoun takes the wh
expression as antecedent are more acceptable if the wh expression conveys the
implication that the construct being questioned is one of a set of individuals
already established in the discourse context, as in questions with which:25
(6.94) Which painting did its owner complain was underpriced?
Secondly, the explicit addition of modifying material to a wh expression substan-
tially increases its viability as an antecedent for some subsequent pronominal {
compare (6.96) with (6.95):
(6.95) What type of car would its owner never part with?
(6.96) ??What would its owner never part with?
Such increased specication of the restrictive predicate for some WH meta-
variable does not need to be made explicit in the wh form itself. Interpreted
against an appropriate context, a question containing an unmodied wh ex-
pression may substantially improve in acceptability. Thus, (6.96) is consider-
ably more acceptable construed against a context in which what is as in (6.97)
construed relative to a contextually restricted set of entities:
(6.97) A: Some cars seem to give endless trouble, but there are a few jewels
that people get irrationally fond of, once they own them.
B: What would its owner never part with?
A: An XJS, to begin with. Then, the old long-wheel-base
Landrovers.
25 Questions with which are commonly said to presuppose that there is an entity of appro-
priate type (see Lappin and Reinhart 1988).
216 Crossover Phenomena
223
224 Quantication Preliminaries
Thus, in (7.1), there is no reversed scope interpretation for the two quantifying
expressions, though in (7.2) and (7.3) such interpretations are marginally pos-
sible. Wide scope interpretations are not possible for either almost nobody in
(7.4) or for few friends of mine in (7.5).
On the other hand, there are the indenites, which may all be construed
as taking arbitrarily wide scope with respect to the environment within which
they are contained, their order relative to other expressions in the string, in
consequence, providing little clue as to how they are to be interpreted:
(7.6) Two of my colleagues wrote an article which a friend of mine
complained about.
(7.7) John is planning to have several parties in which we all celebrate the
work of two Dutch colleagues.
(7.8) John went to see a lm and in it many friends of mine were portrayed
as idiots.
(7.9) John was visiting his mother while two painters I know were stripping
his house.
It has been suggested that the very wide scope eects associated with indenite
determiners are evidence of ambiguity between a sense in which the indenite
acts as a referring expression, its so-called specic use, uniquely picking out
some entity which the speaker has in mind (Fodor and Sag 1982 and many
authors following them) and a regular existentially quantied term, and that
it is the uses of indenite noun phrases as names which give rise to the ap-
parently unconstrained process of quantier construal. However, it has been
demonstrated by a number of people (see Farkas 1981, Cormack and Kempson
1991, Abusch 1994, Reinhart 1997) that intermediate interpretations are avail-
able which cannot be explained by any such supposed discrete referring term. In
these intermediate cases, the indenite can be construed as having wide scope
with respect to the relative clause, but narrow scope with respect to some higher
quantifying expression elsewhere in the clause:4
(7.10) Every professor rewarded every student who read a book he had
recommended.
(7.11) Each student has to come up with three arguments that show that
some condition proposed by Chomsky is wrong.
The assumption, furthermore, that the process of quantier construal can be
analysed as a process of long-distance dependency, in the same terms as move-
ment, feature passing or assumption discharge, faces the additional puzzle that
the indenite determiners do not display any of the constraints associated with
4 Cormack and Kempson (1991) argue that the apparent implication that the indenite is
construed referentially is never more than a conversational implicature.
Scope Eects and Indenites 227
long-distance dependency; (7.6){(7.9) demonstrate this, with the indenite be-
ing inside a relative clause in (7.6) and (7.7), inside the second of two conjuncts
in (7.8), and inside an adverbial clause in (7.9).
If, however, we approach the phenomenon of quantier scope from the per-
spective of incrementally establishing a logical form on a left{right basis, a
dierent range of solutions becomes expressible. Indenites dier from other
quantifying expressions in being able to freely depend on a distinct term of suit-
able type, a freedom redolent of anaphora. A noun phrase { that is determiner
and nominal together { project an annotation on a node of type e; and these
actions may include the projection of constraints on the eventual scope relations
of the form t1 < t2, `term t1 has scope over term t2'. Such scope statements
have to be collected at root node (requiring a formula of type t), where they
must fully determine the semantic evaluation of the resulting formula. Inde-
nite determiners, for example, can be analysed as imposing a constraint that one
argument of the scope relation is a meta-variable, its value to be identied as a
matter of pragmatic choice. Bearing in mind that every propositional formula
is of the form F o(Si : ), with Si a term denoting the index of evaluation for
, we can analyse indenite noun phrases as taking narrow scope with respect
to that term Si when they take wide scope with respect to other quantifying
expressions, if there are any.5 This enables us to say that indenites always
take narrow scope with respect to some term, that term decided by a prag-
matic choice. In (7.12) there is only one choice, this being that the indenite is
interpreted relative to Si, the term representing the index of evaluation:
(7.12) A student fainted.
Notice that, on this view, scopal relations between terms are projected directly
as relations between terms annotating the structure, and not as a relation be-
tween nodes in the structure.
The advantage of this view is that it classies pronoun antecedent construal
and choice of scope construal for indenites together, as involving a pragmatic
choice. Like anaphoric expressions, the scope of indenites is, as we've seen,
not sensitive to island restrictions. Nevertheless, like the construal of anaphoric
expressions, there is strong preference for the scope of the indenite to be es-
tablished on the basis of linear order. Thus mixed quantication sentences are
reported to be ambiguous with two scope-distinguished interpretations if they
contain an indenite following some other quantifying expression:6
5 Expressing the apparent independent scope readings of indenites in these terms is ap-
proximate only, as there are some construals in which the indenite has to be construed as
taking narrow scope with respect to the index, characteristically selected as the second ar-
gument of the T e relation constructed from tense. Pending an explicit account of tense and
modality, we leave this detail on one side. See Farkas 1997 and Kempson 1996 for argu-
ments that construal of indenites may be dependent on temporally sorted variables used to
represent construal of tense/mood.
6 The only exceptions are examples like a certain which may force a wide scope construal
of the indenite (see section 7.3.2.3).
228 Quantication Preliminaries
will be formed using variable binding term-operators. For instance, the phrase
some man will be represented by
(; x; Man(x)) 2 DF o:
This quantied NP can be completely described by xing four parameters:
1. The Binder (e.g. ) indicates the mode of quantication (i.e. existential).
2. The Variable (e.g. x) indicates the variable being bound by the binder.
3. The Restrictor (e.g. Man (x)) indicates the binding domain of the vari-
able. By their restrictors, objects are introduced in a discourse; and pred-
ications on these introduced objects are then constructed from items more
standardly seen as falling within the nuclear scope of the determiner.
These three features exhaustively specify quantied NP denotations in isolation.
However, to characterize NPs as occurring within a sentence, a scope statement
is required.
4. A Scope statement is an atomic DU proposition, using the scope relation
< DF o;e DF o;e , of the form
x<y
stating, in this case, that the quantier binding x has scope over the
quantier binding y.
Once these four features for quantied NP representations are xed, the se-
mantic, i.e. truth-functional, interpretation of a clause projection is fully deter-
mined. With a given Formula value of type t,
= [(1 ; x1 ; 1)=x1 : : : (n; xn ; n)=xn ]
where the (i ; xi ; i ) displayed are the projections of all quantied NPs occurring
in a clause, each with a unique variable xi , we can associate a strict partial order:
B = hB; <B i
where B fx1; : : : ; xn g and <B B B is an irre
exive, transitive relation on
B. Such an ordering re
ects a choice of scope constraints on the terms occurring
in . Consider a basic mixed quantication example such as (7.22), to which
we shall return several times during the course of the chapter:
(7.22) Every dog ate a biscuit.
On the interpretation re
ecting the order of the words in the string, we have
the pair:
(B; );
Quantication 233
where
= Ate((; x; Dog(x))(; y; Biscuit(y)))
<B = fhx; yig
giving rise to the same truth-functional interpretation as
8x(Dog(x) ! 9y(Biscuit(y) ^ Ate(x; y))):
The scope relation x <B y; is interpreted to mean that the quantier binding
x has scope over the quantier binding y. So, if we have the reversed scope
constraint
= Ate((; x; Dog(x))(; y; Biscuit(y)))
<B = fhy; xig,
then this has the interpretation
9y(Biscuit(y) ^ 8x(Dog(x) ! Ate(x; y))):
The pairs (B; ) used in this example are of a special kind in two ways:
all terms involved in occur in some scope relation,
the terms are (strictly) linearly ordered by the scope relations (in this
case, trivially so, because there are only two terms).
In this chapter our discussion of quantication will restrict itself to sentences
exhibiting this special structure. All formulae of rst-order predicate logic (and
non-branching quantication in general) give rise to pairs (B; ) of this special
kind.
So, our analysis of quantication requires that declarative units of type t
with formula = S : for S some temporal index, have a label giving a scope
sequence:
h: : : ; hS; x1 ; : : : ; xn i; : : : ; ti :
where the ordering S <B x1 <B : : : <B xn contains at least all bound variables
in plus the temporal index S . We now add a binary relation symbol `<' to
the DU alphabet such that xi < xj is an atomic DU formula for all xi ; xj 2
D [ MV , and we set:
Denition 12 (Truth Denition for Scope Statements)
Æ T [h: : : ; hS; x1 ; : : : xn i; : : : ; ti : ] j= xi < xj i
xi <B xj , or
= : : : Qxi : : : Q0 xj
, Qxi and Q0xj are unique with respect to
the variables they bind, and xi ; xj occur free in .
This disjunctive condition embodies the fact that x has scope over y, either
if x <B y is explicitly recorded in the scope label, or the quantier binding x
234 Quantication Preliminaries
in has scope over the quantier binding y in . We will see that the latter
condition makes the statement x < y persistent under restructuring.
Having parsed the string which is (7.22), the decorations of the root node of
the resulting tree description may form the set
fT n(0); T e(S ); F o(Ate((; y; Biscuit(y)))(; x; Dog(x))); T y(t); : : : g [
f(S < x; x < y)g
That is, the scope relations, however projected, appear as atomic propositions
involving the bound variables on the root node. These propositions have been
collected in the course of the parse. Because every term will have been as-
signed a distinct variable, we can identify a term with its variable. The scope
propositions can now be used to algorithmically update the Formula value,
Ate(; x; Dog(x))(; y; Biscuit(y)), into the more familiar shape:
fT n(0); F o(S : 8x(Dog(x) ! 9y(Biscuit(y) ^ Ate(x; y)))); T y(t); : : : g:
In this update the quantier binding x indeed has scope over the quantier
binding y.14
We will now discuss the various components of our analysis in more detail.
7.3.1 Quantied NPs
In this section, we discuss the various sub-parts of noun phrase terms, and their
contribution to the interpretation of sentence-sized units. The structures which
we successively introduce are summarized by the display in gure 7.1, to be
presented in turn. As a way of talking about appropriate constructs, we will
dene some monadic predicates, abbreviating modal DU formulae evaluated at
a type e node:
NP
[h: : : ; ei : APL(P : ; (x; Man(x)))]
DET NOM
[h: : : cn ! ei : P : ] [h: : : ; cni : (x; Man(x))]
VAR RESTR
[h: : : ei : x] [h: : : (e ! cn)i : Z (Z; Man(Z ))]
IF f?T y(e)g
THEN put(Indef(+)); make(h#1i); go(h#1i); put(F o(P (; P )); T y(cn ! e));
go(h"1 i); make(h#0 i); go(h#0 i); put(?T y (cn))
ELSE ABORT
In addition, singular indenites, like all other indenites, are subject to a par-
ticular form of scope statement, to which we return shortly. Figure 7.2 gives
the general shape of the actions undertaken by this determiner.
fT n(a); ?T y(e); }g
fh"1 iT n(a); T y(cn ! e); F o(P (; P ))g fh"0 iT n(a); ?T y(cn); }g
Figure 7.2: The action projected by the indenite determiner a. The appearance
of `+' in the label sequence of the top node is a value on the `indeniteness'
feature.
The denite determiners, on the other hand, are treated like pronouns (of
a more expressive variety); they lift the variable (of type e) out of the cn and
turn it into a restricted meta-variable. For this meta-variable an antecedent to
substitute has to be found:
F o(X U (X )[U= (X )])
r l
where l and r are the left and right projection function respectively.15 For
instance, if man projects (x; Man(x)), then the man projects, by function ap-
plication, a Formula containing a meta-variable Uman(U):
APL(X U (X )[U= (X )] ; (x; Man(x))) = UMan(U) :
r l
ELSE ABORT
(where the action freshput() is explained below, it is as the put() action but
introduces a fresh variable). The eect of this denition is that denite noun
phrases such as the man are analysed as anaphoric expressions, with the meta-
variable U of UMan replaced by a pragmatic substitution process. So in the
text John left: the man was clearly tired, we construct for the second clause a
tree whose root node is annotated with F o(T ired(UMan )). This variable U is
replaced by John, where the truth of Man16(John) entails that T ired(JohnMan )
is true if and only if T ired(John) is true.
On this analysis, the problematic uniqueness presupposition attributed to
denite noun phrases (see chapter 1) is not incorporated as a denotational prop-
erty of the term to be constructed, but is a pragmatic process of substituting
some value for the meta-variable, as in the case of pronouns.17
7.3.1.2 Common Nouns
The common noun man contributes the nominal variable and the restrictor as
sub-terms to the nominal :
F o(U; Man(U)) T y(cn)
The lexical entry has as condition a requirement for nominal type cn, and its
action is to introduce an argument daughter with nominal variable of type e:
F o(U) T y(e)
and a function daughter with the restrictor:
F o(X (X; Man(X ))) T y(e ! cn)
Figure 7.3 displays the actions projected by a common noun. The variable of
type e projected by the common noun is passed to the nominal node, its mother.
16 This property is shared with that, this, these, and those.
17 There are two further uses of denite NPs: terms that refer to institutionalized terms
such as (i) and fact that constructions as in (ii), but both of these require separate denition:
(i) You need to see the doctor.
(ii) The fact that I'm tired is irrelevant.
The rst type is idiomatic. The second we suggest, in the light of (iii) and (iv), is a form of
expletive:
(iii) The fact is that I'm tired.
(iv) The received opinion was that I should retire.
238 Quantication Preliminaries
fT n(a); ?T y(e)g
fh"0 ih"0 iT n(a); T y(e); F o(U)g fh"1 ih"0 iT n(a); T y(e ! cn);
F o(X (X; Man(X )))g
What happens with the variable present at the nominal node is then determined
by the determiner annotating its sister. As we have seen, the determiner can
either construct an object at the type e node in which the variable is bound, or
it can pass on the variable to the type e node. This variable can enter into the
nuclear scope by anaphoric means, or it can be copied to a LINK ed structure
to prepare for a restrictive relative. Given such a linkage to the variable node,
the Formula value at the common noun node can be updated to create an
extended restrictor.
The variable introduced has to be fresh. We will use a rather primitive
device to achieve this: we will have an action freshput(x; ) which checks a
nite sequence of variables one by one and uses the rst xi not occurring in
the tree to execute put([xi =x]). Obviously, the number of variables has to be
chosen so as to be large enough to accommodate the trees under consideration:
freshput(x; F o()) =
IF f" # F o(x1 )g
THEN IF f"# F o(x2 )g
THEN : : :
: : : IF f"# F o(xn )g
THEN ABORT
ELSE put(F o([xn =x]))
:::
ELSE put(F o([x2 =x]))
ELSE put(F o([x1 =x]))
Notice the variable x is here a name not ranging over F o values: the boldface `x'
does that. This action is now used in the macro projected by common nouns.
Quantication 239
Common noun M
IF f?T y(cn)g
THEN make(h#0i); go(h#0i); freshput(x; F o(x)); put(T y(e));
go(h"0 i); make(h#1 i); go(h#1 i);
put(F o(X (X; M (X ))); T y (e ! cn)); go(h"1 i);
go(h"1 i); hfV AR(x)g; put(?SC (x)); ABORTi; go(h#0 i)
ELSE ABORT
So a common noun creates an argument daughter where it deposits a fresh vari-
able of type e; it creates a function daughter and annotates it with the restrictor;
and nally, it annotates the type e node, immediately dominating it with a re-
quirement for a Scope Statement hfV AR(x)g; put(?SC (x)); ABORTi involving
the nominal variable.18 The next paragraph will explain this requirement.
7.3.2 Scope
As has been discussed, we will render scope relations between projections of noun
phrases as scope statements of the form x < y expressing that the term involving
variable x has scope over the term involving variable y. These statements will
annotate top nodes in the tree { that is, nodes annotated by T y(t) propositions.
We will dene a monadic `SCope' predicate SC , where SC () indicates that
the Formula value annotating the pointed node is involved in scope relations.
Denition 13 The monadic predicate SC is dened by
SC () df NP () ^ " (T y(t) ^ 9y(( < y) _ (y < ))):
As expressed by this denition, SC () holds at a node if the Fo value is of
type e and this value is involved in a scope relation annotating some dominating
node of type t.
Above, we have seen that a common noun projects a requirement ?SC (x) for
the nominal variable x at the top node of the tree projected by the whole noun
phrase. So, in order to end up with tree without requirements, the requirement
for a scope statement has to be fullled by an action annotating some appropriate
node with such scope statement. These kind of actions will be pragmatic ones,
i.e. actions assigning scope information will be elements of the set P .
Because of the left-to-right nature of the parse, the actual implementation
of the relative scope relations of the various NPs introduced in the sentence has
to be deferred until the sentence has been parsed through. Determining the
scope structure of the eventual representation takes place in two stages. First,
the string is parsed, and the representation is annotated with scope statements.
In the second stage, these statements are used to construct a more familiar
quantication structure.
18 The notation `h; ;
i' indicates `IF , THEN , ELSE
'.
240 Quantication Preliminaries
19 The determiner xes the kind of scope statement through contributing Indef( =+) to
the type e node, and the common noun contributes the variable. So the type e node is
the rst where all relevant information for scoping is present. This is why scope cannot
be determined by the lexicon. On the other hand, if we allow the determiner to dom-
inate its variable in the way the nominal does, i.e. have a tree description of the form:
fT n(a); F o(P (; P [x=l(P )])); T y(cn ! e)g,
fh"0 iT n(a); F o(x); T y(e)g,
fh"1 iT n(a); F o(yP (; P [y=l (P )])); T y(e ! (cn ! e))g;
then all the information for scoping is xed at the level of the determiner; so it can be pro-
jected by the lexicon, and we do not need the intermediate step via the type e node (with extra
requirements). Such a variable would need a tree node of its own on this alternative analysis,
for this is the only way in which we can force its freshness { note the discrete variable for
both determiner and nominal.
Quantication 241
Scope action for indenites:
IF fIndef(+); ?SC (x); ?T y(e)g
THEN gofirst"(?T y(t));
freshput( (U; (U < x, )
?9y(DOM (y) ^ y < x ^ 8z(y < z ! x < z))))
ELSE ABORT
Notice that, though the selection of some meta-variable U for an indenite to
depend on is free, the choice domain consists of the type e Formula values
satisfying DOM (), and so has to be selected from the terms introduced as part
of the interpretation of the string, and not, like anaphora, more freely.
Scope action for non-indenites:
IF fIndef( ); ?SC (x); ?T y(e)g
THEN IF fDOM + (y); 9z(DOM + (z) ^ y < z)g
THEN ABORT
ELSE put(y < x)
ELSE ABORT
Notice that with respect to scope, non-indenites are blind to the variables
contributed by the indenites (this is eected by using the DOM + predicate).
7.3.2.2 Evaluation of the Scope Statements
After the tree has been evaluated to the top node, this node is annotated with
F o()^T y(t) plus scope statements. These statements can be used to iteratively
(and algorithmically) restructure the formula to give it a form that is more
familiar to us. The restructuring is guided by the linear, connected and maximal
S < x1 < : : : < xn chains in the set of scope statements. As we have assumed,
for every formula, there is only one such sequence. The end result will be a tree
with a top node annotated by F o() such that the semantical interpretation
of re
ects the desired scope relations between the quantied NPs. The basic
restructuring move is of the form:
Denition 14 (Q-Evaluation Rule) Given a formula in which the terms
(1x1 1)A; : : : ;
(nxn n) occur, the Q-evaluation rule has the form
h: : : ; hS; x1 : : : ; xn i; : : : ; ti : [xn n =xn ]
h: : : ; hS; x1 : : : ; xn 1 i; : : : ; ti : f x ();
n n n
where for x occurring free in and S a (temporal) index, the values fx (),
for 2 f; ; Qg, and fS () are dened by:
fx () = 8x( ! )
fx () = 9x( ^ )
fQx () = Qx( )()
242 Quantication Preliminaries
fS () = (S : )
So the Q-evaluation rule consumes one variable, the last, from the scope se-
quence in order to restructure the formula. Iteration of this procedure will lead
to S being the only variable left. This variable is then xed to the front of
the formula. We stipulate, in fact, the following incremental relation between
atomic formulae (see remark 6 of chapter 9)
[x =x] F o fx ():
These incremental updates lie at the basis of the following Quantier Rules:
Q1 ffffTTn(na()a;)T; Ty(yt()t;)S; S<<xx;1:;:::: :; x; xn 1<<xxn ; ;FFoo((8[x n =xn ]); : : : ; }g : : : g
xn ( ! )); : : : ; }g : : : g
1 n 2 n 1
and
Q2 ffffTTnn(a(a);)T; Tyy(t()t;)S; S<<xx1; :; :: :: :; x; xn 1<<xxn ; F; Fo(o(9[x n =xn ]); : : : ; }g : : : g
1 n 2 n 1 x n ( ^ )); : : : ; }g : : : g
and
Q3 ff ffT n(a); T y(t); S < x1 ; : : : ; xn 1 < xn ; F o([Qxn =xn ]); : : : ; }g : : : g
T n(a); T y(t); S < x1 ; : : : ; xn 2 < xn 1 ; F o(Qxn ( )()); : : : ; }g : : : g
and
ffT n(a); T y(t); hS i; F o(); : : : ; }g : : : g
Q4 ffT n(a); T y(t); h i; F o(S : ); : : : ; }g : : : g
These rules tell us now explicitly how to restructure F o values in order to satisfy
scoping requirements. Again, the interpretation of (7.22) as:
hS; y; xi F o(Ate((; x; Dog(x)); (; y; Biscuit(y))))
can be restructured to
hS; yi F o(8x(Dog(x) ! Ate(x; (; y; Biscuit(y)))))
by Q1, and this again can give rise to
hS i F o(9y(Biscuit(y) ^ 8x(Dog(x) ! Ate(x; y))))
by Q2. This, nally, leads by Q4 to
h i F o(S : 9y(Biscuit(y) ^ 8x(Dog(x) ! Ate(x; y))))
Notice that, by denition 12, the statement y < x remains true under the
restructuring, because removal of y <B x from the scope relation is compensated
by the fact that now the quantier binding y has scope over the quantier
binding x.
This unravelling process applies to yield bound variable interpretations of
pronouns, as in the sentence
(7.23) Every dogi indicated hei was hungry.
Quantication 243
Encountering the pronominal he in a parse through this sentence, an antecedent
has to be chosen and substituted for the variable U contributed by he. In this
case, the antecedent has been the term (; x; Dog(x)) as it has been constructed.
Having parsed the last word of the sentence, the decoration that results at the
top node (ignoring tense) is
fT n(0); T y(t); F o(Indicate((; x; Dog(x)); Hungry((; x; Dog(x)))); : : : g
The rule Q1 then works as it should in giving
8x(Dog(x) ! indicate(x; hungry(x)))
Compare this to (7.24) under the interpretation indicated:
(7.24) Every dogi who indicated hei was hungry, : : :
In (7.24), the pronominal is construed to give rise to a restrictive relative. The
substitution of the pronoun's meta-variable by the nominal variable x will ensure
that the variable x, introduced by the common noun dog, is linked to the top
node of the LINK ed structure annotated by
fT n(a); T y(t); F o(Indicate(x; Hungry(x))); : : : g:
By the incorporation process of chapter 4, this creates a complex nominal:
fT n(a); T y(cn); F o(x; Dog(x) ^ Indicate(x; Hungry(x))); : : : g:
Restructuring by rule Q1 then results in the desired
8x((Dog(x) ^ Indicate(x; Hungry(x))) ! : : : ):
These various rules will then interact to capture the fact that if a pronoun
in a relative within a quantifying expression is construed as identied with
some other term external to that expression as in (7.25), the only possible
interpretation is one in which the indenite is construed as taking narrow scope
with respect to that other term, with the unravelling process constrained by the
variable structure inside the Fo value:
(7.25) Every dog choked on a biscuit he ate.
Example (7.26) requires a dierent treatment:
(7.26) Every dogi ate a biscuit. Hei indicated hei was hungry.
Here the term (; x; Dog(x)) that is introduced on encountering the noun phrase
every dog, needs to be preserved (in a sense to be made clear in section 7.3.3),
in order to serve as an antecedent for the pronominal in the second sentence.
But this initial sketch at least indicates how such a term-based analysis can
yield interpretations of pronouns in both the restrictor and the nuclear scope of
a quantifying expression, as in (7.23){(7.25).
244 Quantication Preliminaries
This process of term construction has been formalized in a logical system (for
discussion see Meyer-Viol 1995), which has the following two rules concerning
the terms:
9x(x) (; x; (x))
(; x; (x)) 8x(x)
Quantication 247
Addition of these two rules to a standard system of (classical) rst-order pred-
icate logic does not add to its classical power.21 Notice that, using these rules
in combination with the standard quantier rule, we have:
f8x(Dog(x) ! 9y(Biscuit(y) ^ Ate(x; y)))g `
fDog(b) ! (Biscuit(ab ) ^ Ate(b; ab )))g:
where the abbreviations b and ab are as dened above.
In our set-up this semantics holds only after the scope relations have been
incorporated into the terms; a semantics for the formula tout court is not den-
able until after we have reached an empty scope sequence. This means that not
all constellations of formula and scope sequences lead to formulae which make
semantic sense. For instance,22
h: : : h i : : : ; ti : Dog(; x; W alk(x))
has no proper semantics, but in a dierent scope environment:
h: : : hxi : : : ; ti : Dog(; x; W alk(x))
the formula can be further restructured to:
h: : : h i : : : ; ti : Dog(a) ^ W alk(a)
where a = (; x; (Dog(x) ^ W alk(x))). Now we have a semantics, because
9x(Dog(x) ^ W alk(x)) $ (Dog(a) ^ W alk(a)):
In order to construct the appropriate terms (b and ab respectively) from the
initial structures, for example (; x; Dog(x)) and (; y; Biscuit(y)) together with
the scope statements, rather than from the predicate logic formula, we give the
following update rules which should be compared to those of denition 14.
Denition 15 (Term Restructuring) For in which the variable x occurs
free
fx () = [a=x] ! [a=x]
where a = x( ! )
fx () = [b=x] ^ [b=x]
where b = x( ^ )
fQx () = ( [c=x])([c=x])
where c = Qx(( )()).
21 It is a so-called conservative extension of rst-order predicate logic.
22 Here we give the display strictly as a node on the tree (with the Type and other labels,
labelling the Formula ), since it is necessary to be able to refer to the absence of a scope
sequence.
248 Quantication Preliminaries
As we have now seen several times, the parse of the string Every dog ate a biscuit
results in a tree the top node of which is annotated by
(S < x < y) F o(Ate((; x; Dog(x)); (; y; Biscuit(y)))); T y(t):
The term-restructuring rules now allow us to generate by the second rule the
structure:
(S < x) F o(Biscuit(a) ^ Ate((; x; Dog(x)); a)); T y(t)
where
a = (; y; (Biscuit(y) ^ Ate((; x; Dog(x)); y)))
Notice that (; x; Dog(x)) is a sub-term of a. This can be seen as dening
(; x; Dog(x)) < a. Then applying the rst rule gives:
to the predicate. The motivation for this was the need to provide an analysis
of indenites as lexically projecting an underspecied scope statement. The
decision has a number of advantages. It makes it possible to provide a basis
from which E-type anaphora can be explained as part of a uniform pragmatic
account. Moreover, the compilation of epsilon terms with complex internal
structure means that these anaphoric eects can be expressed without any ad-
ditional statement. It also provides a natural basis for characterizing lexical id-
iosyncrasies constraining scope which are quite unstatable in terms of a general
storage mechanism. More generally, it provides a relatively natural re
ection
of the way in which choices, made on a left{right basis, aect scope of terms
within some logical formula. And the phenomenon, as so described, provides a
further instance where pragmatic actions feed into an algorithmic process { here
that of establishing the logical representation of the content of a string. This
account thus gives one more example of a structural process of interpretation
that follows the left{right
ow of the parsing process and involves interaction
between computational and pragmatic actions.
8
Re
ections on Language
Design
8.1 The Overall Perspective
Throughout the previous chapters, we have been outlining and arguing for a
formal model of language which purports to be at one and the same time a
model of an incremental left{right process of interpretation for natural language
strings, yet also a basis for explaining structural properties of language, and
hence a grammar formalism. We now look back on what we have done in order
to be able to re
ect on the relevance of the results for general linguistic theory.
Our starting point in this endeavour, in chapter 1, was to set out the prob-
lems for semantic accounts of anaphora by way of background motivation for ex-
ploring in some detail a representational account of the process of interpretation.
In case after case of pronoun{antecedent pairing, we found that characteriza-
tion of pronominal construal through a semantic characterization of the relation
between the expression and the object denoted was far from straightforward.
What appeared to be needed was some representational form of explanation,
analysing a pronoun as projecting a skeletal indication of some resulting inter-
pretation to be replaced by a term selected either from what is made available
in the build-up of interpretation of the sentence under evaluation or, more gen-
erally, from implications drawn from the previous discourse.
In chapters 2{3 we dened a modular system which, across a left{right
sequence of words, induces a structure from which some lambda term is compiled
representing an interpretation of the string. The progressive update of partial
tree structures re
ects three properties central to natural language processing:
(i) The parsing process is goal-directed, the goal being to establish some
logical form as representation of a propositional content for some sentence.
(ii) The information needed to establish such interpretation is progressively
accumulated across the left{right sequence until the goal is achieved.
253
254 Re
ections on Language Design
another. With this extension, we then dened the process of constructing se-
quences of tree structures. Relative clauses were taken to project a LINK ed
tree independent of the primary structure except in so far as there was an im-
posed requirement on the development of the LINK ed tree from its top node
that the formula to result at that node must have as a sub-term a copy of the
formula annotating the node from which this otherwise independent tree was
developed. Given the dynamics of the system already developed for individual
tree growth, this constraint can be satised in one of a number of ways in the
emergent LINK ed structure.
This interaction of transfer of information from one tree to another and the
dynamics of tree growth for an individual tree enabled us to capture variations
both within a single language and across languages. The rst major result was
that the pattern of behaviour of pronouns in languages such as Arabic, where
resumptive pronouns are obligatory in non-subject positions in relative clauses,
was naturally achieved by dening the LINK transition for that language in
terms of an imposed requirement on the top node of the LINK ed structure for
a copy in that structure of the head formula, and not as some annotation on an
unxed node (as in English). The resulting characterization of the Arabic data,
both for relative clauses and for clitic left-dislocation structures, was achieved
without any structure-specic stipulation about the pronoun itself. It then
emerged that the concept of LINK ed structure applied equally to head-initial
and head-nal relative clauses, with minor variation in the rule formulation.
Within this general picture, the so-called head-internal relatives are also un-
problematic (unlike in other frameworks): in these relatives, the term in the
LINK ed structure is constructed in the normal way, duly providing a value for
the subsequent head formula by a simple process of copying. The result overall
was a unifying relative clause typology of considerably greater generality than
other analyses of relatives.2
In chapter 5 this style of analysis was applied to what, at rst glance, seems
a totally dierent type of phenomenon { the dierent forms of wh expressions
in questions, spanning wh-initial expressions, wh-in-situ expressions, and the
supercially very dierent phenomenon of wh -expletives displayed in some lan-
guages. In other frameworks, wh expletives constitute a scoping problem, the
expletive element `stretching' the scope for the full form of the wh expression
which follows. In these frameworks, the analysis requires additional stipula-
tion, though advocates of dierent analyses disagree as to how much additional
stipulation is needed. In the present system, no extension of the framework is
needed. It is simply that a lexical expression can be anticipatory for a second
expression by providing the annotations projected by the second expression as
a requirement. Indeed, the concept of having lexical expressions in a language
whose role is anticipatory is entirely natural within this framework, opening the
way for a general theory of expletives encompassing categories of all types (see
Perrett 2000 for an analysis of verb-chaining phenomena in Hadiyya in terms
2 The only analysis which comes close to the breadth of this characterization is that of
Kayne 1994; see chapter 4, n.42.
The Overall Perspective 257
of an expletive form of the tense marking). The analysis of wh expletives was
broadly successful at the descriptive level, covering the same data as other anal-
yses, but furthermore incorporating other data (specically the non-was forms
of expletive), an integration which is new. As in the case of the analysis of
relatives, the result is a typology incorporating the major type of wh question
words and their position, according to whether they project an annotation to
an unxed or a xed node, or a requirement on subsequent tree development.
Crossover phenomena, which were the subject of chapter 6, involved the
interaction of all the analyses so far set out { construal of pronouns, relatives
and wh questions. Its primary interest lay in the demonstration of how these
analyses, independently set up and motivated in previous chapters, could be
seen to yield an integrative account of what had previously been analysed as
a heterogeneous set of phenomena. The distinguishing property of our expla-
nation is its critical reference to the process of interpretation { the possible
construal of pronouns depended on what information is available at interme-
diate points in the interpretation process and what consequences follow from
any particular selection of a value. This was the heart of the analysis of the
crossover phenomena, the dierences between strong and weak crossover in in-
dividual languages turning on the details of what was made available by the
LINK transition, the node within the structure annotated by the wh expres-
sion or pronoun, and interaction with general transition rules. The additional
novelty of the account of weak crossover phenomena was its exploration of the
consequence of allowing interaction between computational and pragmatic ac-
tions. Dening pronoun construal as substitution of some arbitrary formula,
without any system-internal constraint on how that selection might be made,
means that the system itself is extremely liberal with respect to the data, for the
unacceptability of interpretation for a given string may be due to considerations
entirely external to the system itself. In the case of weak crossover data, we
argued that the lack of robustness of weak crossover eects in questions was a
consequence of such system-external pragmatic constraints, suggesting that the
phenomenon is not due to a structural restriction at all.
In chapter 7 we looked at quantication as a challenge to our general commit-
ment to a left{right form of incrementality of build-up of interpretation, for the
global nature of quantier scope seems to deny the possibility of sustaining an
incremental approach to natural language understanding. However, we made a
separation between scope statements and rules for determining the denotational
content of quantier expressions, and then dened a way in which constraints
of relative scope statements can be gathered in an incremental way as the tree
and its annotations are set out. This characterization makes possible the artic-
ulation of lexical constraints on scope determination which are inexpressible in
any system in which the scopal properties of quantifying expressions are dened
through a general process such as quantier movement or quantier storage. Fi-
nally, we demonstrated, albeit brie
y, how this account of quantier construal
could be used to project bound variable and E-type forms of anaphora con-
strual. The particular signicance of this account of quantication is that it
provided another case where the pragmatic action of Substitution must feed an
258 Re
ections on Language Design
algorithmic process, that of determining the nal logical form projected by the
string, hence requiring interaction of computational and pragmatic actions.
Once this way of looking at natural language data becomes familiar, whole
sets of phenomena suggest analyses in similar terms. We might analyse verb
raising (the verb-second phenomenon in Germanic languages) in terms of pro-
jecting some locally unxed node to be annotated by the verb. Clitic raising
also, we might analyse as the projection of an unxed node, this time for the
pronoun to annotate. Clitic doubling might be analysed as projection of a re-
quirement by the pronoun, rather than an annotation. And so on. Many of
these analyses have been only hinted at here, but, seen as dierent manifes-
tations of the progressive update of structural information, not a single one
of these phenomena should be taken as an indication of `imperfection' in an
otherwise `perfect' system (Chomsky 1995). To the contrary, they display the
natural language system for what it is { a vehicle for updating partial descrip-
tions through from some initial goal to a logical form in satisfaction of that goal.
In short, they re
ect the dynamics intrinsic to the parsing process.
Overall, it has to be conceded that what we have achieved in formal terms is
rather little; for we have barely provided enough detail to see how a full fragment
could even in principle be dened. In particular, on the formal side, we have
not provided a full characterization of control of the construction process and
pointer movement. And from the data point of view, several important types
of phenomena are missing: we have done no more than sketch an account of
tense construal, there is no characterization of connectives, hence of complex
constructions such as conditionals; and there is no characterization of non-nite
constructions or any of the so-called A-chain phenomena. As a construction of
the interpretation process too, the model provided dened is also not more than
partial.
The decision to apply the framework by individual sketches was deliber-
ate, however, for each of the problems we chose to address in some detail is a
puzzle, given current assumptions. To our knowledge, there has been to date
no attempt within any of the standard frameworks to provide a general ac-
count of pronoun construal which encompasses discourse construal,3 interaction
with long-distance dependency phenomena, and resumptive uses. The exis-
tence of head-internal relative clauses is acutely problematic for all standard
accounts of relative clauses, dened as they are in terms of the head (or its
associated relativizing operator) binding a position within the relative. W h
expletive phenomena pose another puzzle for syntactic theories, highly prob-
lematic for Minimalist assumptions. Weak crossover phenomena have remained
an acknowledged `mystery' since Postal (1993), and no attempt has been made
to relate these data to the phenomenon of strong crossover (with the exception
of Dowty (1993) working from categorial grammar assumptions). Yet, relative
to these new assumptions, the data have emerged as a consequence of artic-
ulating the dynamics of the goal-directed interpretation process, with natural
3 This, despite the fact that the changing interaction between anaphora and grammatical
processes is a well-known basis for language change. This is a well-known and traditionally
much-studied phenomenon, re-emphasized by Givon (1976).
Underspecication and the Formal Language Metaphor 259
cross-language typologies determining the limits of variation. The individual
sketches gathered together thus display the potential for new insights to be
gained by adopting the very dierent perspective proposed here { that a for-
mal characterization of the left-to-right process of building up representations
of interpretation over a sequence of words itself provides the basis for explaining
syntactic properties of such strings without need of any further level of repre-
sentation. Hence, coming round full circle to the outset of the book, we take the
model we are articulating to constitute a formal reconstruction of the common-
sense view that knowing a language involves knowing how to use it to recover
some intended interpretation.
8.2 Underspecication and the Formal
Language Metaphor
It is now time to take a step back and, on the strength of these empirical results,
to reconsider some of the general issues raised initially in chapter 1. There are
three major ways in which this framework is novel (though with respect to the
rst of these, see Hausser 1989, 1999).
(i) Structure is projected on a left{right basis across the sequence of words.
(ii) Structural concepts of underspecication are licensed, so that not all tree
descriptions constitute complete trees.
(iii) The concept of syntax is shifted to incorporate the dynamics of progres-
sively building up incomplete structures.
The rst property is self-evident: the projection of a logical form is not on a
purely bottom-up basis, but through applications of rules for unfolding a tree in
a goal-directed, top-down way as driven by the sequence of words. The second
and third properties are two sides of the same coin, the one a consequence of
the other. With a concept of underspecication incorporated into syntactic
dscription, the denition of a dynamic tree-update process becomes essential.
Unlike in other frameworks, the characterization of underspecication of in-
terpretation for a natural language expression and its resolution has not been
articulated within a system of semantic evaluation of some antecedently dened
structure. To the contrary, in our system, the diversity of construal of natu-
ral language expressions in context results from a constructive term-building
process for which natural language expressions provide input procedures. It
is not semantics, then, which is taken as the basis of the unfolding dynamics
of natural language understanding. It is syntax that has been made dynamic,
with the remit of syntax redened as covering the projection of sequences of
partial tree structures. Phenomena are described not in terms of some complete
structure or structures, but in terms of transitions across partial and growing
tree descriptions. Accordingly, the explanation of anaphora and the diversity
of interpretation of natural language expressions in context is now analysed in
260 Re
ections on Language Design
well-formedness which is more liberal with respect to the set of strings licensed,
since whenever an informant characterizes a string as acceptable in virtue of
there being a string{interpretation pairing, we would characterize such a string
as well-formed (unless there was strong evidence to the contrary). So, to take
one type of example prominent in chapter 6, we would characterize (8.1) as well-
formed, on the grounds that it is acceptable under certain circumstances:8
(8.1) The man that Sue said he needs a holiday is chairing the meeting.
In short, the relationship between theory and data is much as in other frame-
works. Despite the fact that we are promoting a grammar formalism which sets
out a dynamic architecture for parse sequences, it is none-the-less possible to
dene a clearcut concept of well-formedness and describe a further concept of
acceptability which re
ects some interaction between a grammar formalism and
pragmatic constraints on retrievability of interpretation.
8.4 Universals and Language Variation
A further re
ection, given a broad Principles and Parameters or Minimalist
perspective, is that the analyses given for any one phenomenon have, disap-
pointingly, invoked dierent forms of rule in each language considered. To take
one example, far from there being a single LINK transition rule capturing what
is common to relative clauses, to the contrary, there may be more than one LINK
transition rule in even a single language (as in English). And although the pair
Introduction (which introduces requirements for daughter nodes) and Prediction
(which licenses the construction of such nodes) might seem good candidates for
a set of universally applicable rules, not all language-particular analyses have
made use of them. Japanese, in particular, was dened as inducing the full se-
quence of constituents from the absolute root by lexical action. And in English,
these rules were used only to introduce the subject and predicate nodes. Even
lexical items of the same syntactic category do not necessarily project the same
sequence of actions. Though transitive verbs, for example, have been assumed
to project a formula of type e ! (e ! t), they vary freely with respect to other
accompanying actions they induce. In some languages they have been dened
to initiate a set of actions from a requirement ?T y(t), as in Arabic (in Japanese
even building rst an unxed node of that requirement), in other languages
from a requirement ?T y(e ! t), as in English; and, as suggested in passing
8 One problem in evaluating judgements of acceptability in borderline cases is that infor-
mants may make dierent judgements according to whether they place themselves in the role
of hearer or in the role of speaker. Examples such as (8.1) are treated as unproblematic in
language use, occurring regularly in conversational English, but nevertheless may be reported
as a form which the informant as speaker might avoid. All such cases would be included by us
in the set of data to be characterized as well-formed, unless to do this would lead to internal
inconsistency in the system, as in the case of agreement mismatches which all informants
agree are unacceptable despite being understandable:
(i) The men is seeing my mother.
Unlike (8.1), examples such as (i) do not occur regularly in conversational English.
Universals and Language Variation 265
in chapter 5, we might dene German verbs as initiating a set of actions from
a condition ?T y(e ! (e ! t)) (presuming on Introduction and Prediction to
induce the full template of structure for a predicate and its arguments). More-
over, the characteristic pattern of lexical entries is a disjunctive specication,
projecting dierent actions in dierent environments.
A follow-up question re
ecting on this rich lexical diversity might well be:
in what sense does the syntactic explanation oer anything more than a het-
erogeneous list of dierent forms across dierent languages? The answer to this
question lies in the constraint that the process of natural language interpre-
tation is by denition goal-directed. The interpretation of a natural language
string involves goal-directed construction of an annotated tree structure, guided
by the sequence of words, the goal being to yield a composite lambda term of
type t constructed out of other such terms (not expressions in the string under
interpretation). The form of that goal, its initial statement as a goal to be
achieved, and the format of the increasingly rich tree descriptions which consti-
tute the transition from the assigned goal to its achievement are all determined
by the particular tree-description language, together with the selected systems
for representing node decorations, being a combination of LOFT, a variant of
the lambda calculus (with term-binding operators) for expressing formulae, and
other labelling devices. To express the dynamics involved in such a sequence of
transitions, the vocabulary for describing node decorations includes both anno-
tations and requirements. Moreover, the decision to adopt a modal logic as the
tree-description language, with h"i, h#i, hLi, hL 1i, h#i, #, h"i, ", hDi, hU i,
means that requirements can be dened which are locally imposed on a certain
node but may nevertheless be fullled by annotations on some node which is
only related globally to the node at which the requirement is introduced. In
addition, individual annotations may be partial, requiring substitution by some
xed value in all such cases to yield a completed tree. Thus, for example, the
language for representing Formula values has meta-variables U, WH, etc., and
the tree-description language has h#i, hDi, hU i, these being devices for express-
ing concepts of incompleteness of annotation along the dimension provided by
the particular label, F o and T n, respectively. Finally, given the gap between
the imposed goal and its achievement, the processes which develop tree struc-
tures are dened so as to impose further requirements, project annotations, and
compile these so that the goal can be attained. The framework, that is, dictates
that natural languages be described as sets of strings interpreted by an incre-
mental left-to-right process monotonically constructing partial decorated trees
to yield a logical form; and the structural properties of a language are dened
in terms of the contribution made to this process by words in the language in
combination with general transition rules. It should be noted that there is no
unrestricted process of type construction, so that there are only a nite number
of Type annotations available for node decoration. In other words, there is no
general process of type inference, as in categorial grammar formalisms.
What varies within the space of possibilities provided by these assumptions
is how the goal imposed on the interpretation of a particular string is achieved
in any one language, and, on this, there are no additional a priori restrictions.
266 Re
ections on Language Design
268
Introduction 269
where for 2 ACT and a formula `based on' AT OM , we have T j= [] if
for every T 0 2 P T : if T < T 0, then T 0 j= .
As usual, Vp can be lifted to assign to elements of P T the formulae of a
language DU , the closure of AT OM under the Boolean operations, quantiers
and modalities, and Va can be lifted to assign to pairs of states in P T the actions
in A(ACT ), the closure of ACT under the regular operations, plus test, by the
standard P DL truth denition.
To t our model in this mould, we have to keep in mind that the actions in our
model are generally to be interpreted as constructions that contribute to a
logical form, and that these forms constitute some `maximum' for these actions.
This motivates the following specialization of the standard P DL model.
Specialization 1: a partially ordered set of states
On the domain P T we introduce a binary relation , the development relation,
where is a (discrete) partial order3 with a unique minimal element T and
a set of maximal elements P T+ such that for all T 2 P T there is T+ 2 P T+ :
T T T+ . The set P T+ represents the objects to construct (e.g. logical
forms), and the elements of P T are partially constructed such elements. The
actions of P DL are now seen as constructions. In other words, for all 2 A,
< :
That is, the states in P T+ are the objects to construct. The actions incremen-
tally construct these states, starting from T .
If we consider a path T : : : T : : : T+ 2 P T+, then the parts of the
maximal states4 under construction are exactly the persistent features; that is,
the formulae such that T j= and T T 0 implies T 0 j= . There may also
be such that T j= which do not persist in this way. Such formulae represent
transient features of states and may function to control actions, but they are
not parts of the objects under construction.5
The maximal elements in P T+ (the set of logical forms) divide the (parsing)
3 An order is discrete if, whenever T T 0 ; T 6= T 0 , then there is a T 00 : T T 00 T 0 such
that for every T 000 such that T T 000 T 00 , we have T 000 = T or T 000 = T 00 : if an element
has a successor, then it has an immediate successor.
4 The parts of the maximal states are trees decorated by declarative units.
5 Partial trees are not the basic elements of our model: the basic elements are the partial
trees plus a pointer; that is, partial trees with one selected node. The elements T 2 P T
now become equivalence classes consisting of collections of all pointer positions over one and
the same tree. On a non-empty domain P P T = fT n j T 2 P T; n 2 T g we place two binary
relations and , where is a (discrete) partial order as above, and an equivalence relation
(between two pointed partial trees which dier only in the pointer position), such that for all
2 A,
< ( [ ):
That is, we consider a P DL model with actions that are incremental with respect to some
partial order on the domain of states. All actions move along , except pointer movements,
which go along . We can identify the elements of P T with the objects over P P T , where
this set is dened by O(P P T ) = f[T n] j T n 2 P P T g. Objects are related through their
elements: [T ] [T 0 ] i for all T 00 2 [T ] there is a T 000 2 [T 0 ] such that T 00 T 000 .
Moreover, if T T 0 , then [T ] [T 0 ] .
270 The Formal Framework
paths in P T starting from T1 into `correct' and `incorrect' ones: correct pathways
are those that have continuations reaching P T+; incorrect pathways are then
the ones that don't. A formalization of this goal-directedness is our second
addition to P DL.
Specialization 2: requirements
The addition of requirements to the P DL model with a partially ordered domain
of states consists in the addition of a (third) function V? besides Vp and Va ,
V? : P T 7! P (DU )
where we set T j=? if 2 V? (T ), which is intended to mean that is `required'
at T (note that this function is not dened by induction on AT OM ).
The function V? satises the following constraint:
T j=?
which implies
T 6j= ,
if T T 0 and T 0 j=?, then 8T 00 (T T 00 T 0 ) T 00 6j= ).
So a requirement for at T persists along until a T 0 is reached satisfying
. Note that may itself be non-persistent or persistent. Moreover, all non-
maximal states T (i.e. T 62 P T+) must have outstanding requirements. This is
what goal-directedness means in our model:
fT 2 P T j V? (T ) = ;g P T+
On the other hand, the minimal state cannot be without requirements. The
transition from state to state must be driven by the need to full requirements:
V? (T ) 6= ;
Apart from an ordering on the P DL states and the addition of a third `valuation'
function, our parsing model specializes the general P DL framework further, in
that the states themselves are structured objects (decorated partial trees), and
the basic actions are not arbitrary, but are specic to the states of our model.
Specialization 3: structured states and specic actions.
Abstractly, the parsing process of a natural language string w can be viewed as
establishing an association between w and a set S of terms (formulae) in some
logical language, the possible interpretations of w. In any parsing framework,
this interpretation is likely to be the result of applying some (set of) basic
operation(s) O to the representations contributed by the words of string w. In
fact, the representation is generally a normal form under these operations.
This means that the association between string w and set of formulae S can be
broken down into associations between the individual words of w and sets of
representations labelling the frontier of a binary tree the top node of which
is labelled by S :
Introduction 271
O(X3 ;O(X1 ; X2 ))
X3 O(X1 ; X2 )
X1 X2
For instance, we may take X1; X2; X3 to be feature structures and the operation
O to be unication, or X1 ; X2 ; X3 to be lambda terms and O Application, or
X1 ; X2 ; X3 to be labelled categorial expressions and O Application: Modus
Ponens, or X1; X2; X3 to be DRS s and O Merging. In all these grammatical
frameworks the construction of terms in a particular parsing process can thus be
represented as the growth of decorated binary trees. In this book we will take
the operation O to be function application in a typed lambda calculus, and the
objects of the parsing process, that is the elements of P T+ , will be terms in this
calculus together with some labels; but it is important to keep in mind that the
choice of the actual representation language is not central to the parsing model
developed here.
So the maximal states in P T+ in our model will take the form of labelled
formulae of a typed lambda calculus. The labels record, among other things,
type information and linguistic features like tense, agreement, etc. A formula of
the lambda calculus will be represented by a decorated binary tree as follows.
The sentence John read a book, represented by the formula Read(John, Some(x,
Book(x))), can be seen as resulting from the term
APL(APL(xy [Read(y )(x)]; APL(P (SomeP ); (x; Book (x))); John))
by -reduction. The obvious tree structure of this term we will exhibit in the
form of a bracketed formula:
[0 [0John] [1[1 xy[read(y)(x)]] [0 [1P (someP ) [0(x; book(x)) ]]]]]
where 0 represents the argument and 1 the function daughter. Tree and decora-
tions can be considered independently by pulling them apart6 (gure 9.1). This
supplies the abstract point of view of the following sections.
[|0 [0 ] [1 [1 ]{z[0 [1 [0 ]]]]]}
T ree
|f00:john, 011:xyread, 0101:{zP (someP ), 0100:(x; bookx)g}.
Decorations
the function, holds at the function daughter, and 1 , the argument, holds at the
argument daughter of n. Now, -reduction APL(1 ; 2 ) = , can be represented
by (instances of) the axiom scheme APL(1 ; 2) ! , relating the propositions
on function and argument daughters of n to propositions at n (see Blackburn
and Meyer-Viol 1994).
The minimal element T in this set-up is a one-node tree decorated by a
requirement for an annotation of a formula of type t { that is, a requirement for
a proposition. The parsing process which maps T eventually to a decorated
tree in P T+ creates on the way decorated partial trees. These will constitute
the domain P T of our model.
Now, we will view parsing as the construction of a (set of) decorated binary
trees, logical representation(s), guided by the input of a natural language string.
This can be modelled as (incremental) growth of a (set of) decorated tree(s).
This is modelled by the relation on the states. In this book we will assume
that this construction takes place while traversing the natural language string
from left to right and in a word-by-word fashion. Every word maps a partial
decorated tree to an extension, a decorated partial tree in which it can be em-
bedded. The input string licenses the growth of a particular collection of trees
in the forest of possible trees. From left to right, every word in the string may
induce a specic extension of the current (set of) partial tree(s) (sentence-initial
John, The), or it may more generally constrain the possible completions of the
current partial tree (John who : : : , John is going to project some argument in
the tree constructed by the relative clause). There are various components to
this growth relation: the growth of tree structure, the growth of tree decorations,
and the disappearance, i.e. fullment, of requirements.
First, a tree structure may be extended by growing new branches, extending
old ones, etc. But the tree structure may also be underspecified: it may be
underdetermined whether some daughter is a function or argument daughter {
thus we may have [n : : : [# : : : ] : : : ] representing the relation of immediate
dominance without having more specic information. Also, the tree may be
underdetermined in the fact that only knowledge about dominance is involved
{ denoted by [n : : : [ : : : ] : : : ]. In such a partial tree, the fact that one node
immediately dominates another one does not (yet) mean that the second one is
a function daughter or an argument daughter of the rst one, and that one node
dominates another one does not mean that these nodes are related by a sequence
of immediate dominance steps: this holds only in the (complete) binary trees in
which it can be embedded.
Secondly, in the course of the traversal of the natural language string the for-
mulae decorating the nodes of the tree under construction may grow in number
and also in information content. That is, they may be updated and, in particu-
lar, variables may be substituted.
Thirdly, the growth of a partial tree is not arbitrary. The string John read
: : : leaves us with the structure
[0 [0 ], [1[1],[0 ]]] f00 : John; 011 : xyRead(y)(x)g:
Declarative Structure 273
This structure is a partial decorated tree that cannot be reduced to a lambda-
free term: at least one argument is still required to be able to reduce the tree to
a term in the representation language. That is, this partial tree restricts the set
of possible completions to the one where an object to read is supplied (see gure
9.2). Here is where the function V? comes into play. On the way to the complete
[|0 [0 ],{z[1 [1], [0]]]} f| 00 : John; 010 : xy[read
{z
(y)(x)]; 011 :?T y(e)g}
T ree Decorations
9.2.1.1 T -structures
The decorated binary trees that will serve as our representations of logical forms
form a subset of the decorated partial trees. These, in turn, are special kinds of
decorated T -structures. In this section we formalize the structural parts of the
decorated structures, i.e. bare T -structures, partial trees and binary trees; the
next section, 9.2.1.2, will concentrate on decorations, and section 9.2.1.3 will
then deal with the structures supplied with decorations.
Denition 16 (T -structures) A T -structure is a tuple
T = hT; 0; 1 ; L; # ; ; D i
where T is a nite non-empty set and all of the i are (possibly empty) binary7
relations on T .
So, a T -structure consists of a non-empty domain on which there are a number
of binary relations dened. Without further ado, these relations do not struc-
ture the domain as a binary tree. This will have to be taken care of by a next
denition.
We let I = f0; 1; L; #; ; Dg, and I (as usual) will be the set of all nite se-
quences of elements of I (including the empty one). For j 2 I and i 2 I we set
ji =j Æ i .
Structural similarity of T -structures is measured by T -morphisms. A function
f mapping T into T 0 is a T -morphism from T -structure T into T -structure T 0
if for all n; m 2 T , for all i 2 I : n i m ) f (n) i f (m). A function f is a
T -isomorphism from structure T to T 0 if f is a bijection from T onto T 0 and
both f and f 1 are T -morphisms.
Denition 1 (Binary Trees) A T -structure T is an (ordered) binary tree if
hT; # i is a directed, non-circular graph such that T has exactly one ele-
ment without # predecessors, and every element with # successors has
exactly two such successors,
0 and 1 are partial functions such that # = (0 [ 1 ) and
(0 \ 1)= ;,
is the re
exive, transitive closure of # .8
A T -structure T is a binary linked tree if T is the disjoint union of a nite set
of binary0 trees such0 that n L m, for m 2 Ti 2 T , implies that n 62 Ti and
n L m for all m 2 Ti . We dene D as follows: n D m i there is a
0 k ! such that n i1 m1 : : : i 1 mk 1 i m, where for all 1 j k:
ij 2 f#; Lg (D may relate nodes from distinct binary trees).
k k
The set of T -structures that are binary linked trees will be denoted by BT .
7 The symbol `' will be used to indicate the end of a denition, remark, etc.
8 We will also nd use for the relations 0 and 1 which are the re
exive and transitive
closure of 0 and 1 respectively.
Declarative Structure 275
Remark 1 Notice that in linked binary trees we have
0 ; 1 #D and L D :
From right to left we have increasingly `specied' tree relations. The most
specic are the 0 and 1 relations. Both these relations are immediate dominance
relations (#), and immediate dominance relations again are dominance relations
() which, in their turn, are D relations, possibly involving more than one tree.
The 0 and 1 daughter of a node will be called the argument and function daughter
respectively. The result of applying the formula annotating the 1 daughter of n
to the formula at the 0 daughter is deposited at n itself.
Apply(Function, Argument)
0 1
Argument Function
For n 2 T , the local tree (global tree) of n is the smallest set T R such that n 2 T R
and if m 2 T R, n j m or m j n, j 2 f0; 1; #; g (j 2 f0; 1; #; ; L; Dg) then
m 2 T R. So the local tree of a node n consists of the binary tree it is a node of.
Now we are set to introduce the basic structural objects of our model: partial
trees. These are the structures underlying the partial logical forms the parser
constructs.
Denition 2 (Partial Trees) The set P T , of partial trees, consists of all T -
structures T such that there is a T -morphism mapping T to an element of BT ,
a binary linked tree.
So a partial tree is a T -structure that can be embedded in a binary linked
tree. We are going to conside the partial trees dynamically as `states' in a
construction process. This construction process will take place at the nodes of
a partial tree; that is, we need a tree together with a pointer indicating some
node.
Denition 3 (Pointed Partial Trees) A pointed partial tree (ppt) T n con-
sists of a pair of a partial tree T 2 P T and a node n 2 T , the domain of T . The
set of all ppt will be denoted by P P T . The System of ppts, PPT , is the tuple
PPT = hP P T; fij i 2 I gi
where T n i T 0 n0 if T = T 0 and n i n0.
Partial trees are related by the basic relation of extension, development or
growth.
Denition 4 (T -extensions) A ppt T 0 n0 T -extends the ppt T n, notation
T n T T 0 n0 , if there is a T -morphism f mapping T to T 0 such that
f (n) = n0 .
Notice that T is a pre-order on P P T (i.e. it is re
exive and transitive). This
relation on pointed partial trees can be lifted to the partial trees; for T ; T 0 2 P T
we set T T T 0 if for all n 2 T there is a n0 2 T 0 such that T n T T 0n0.
276 The Formal Framework
D E B C
It can also mean that adds to T a completely unconnected partial tree,
T0
creating, in eect, a set of two disjoint trees. Moreover, in T 0 nodes of T may
have been unied (the f0 -images of two distinct elements of T coincide, under
T -morphism f : T 7! T ).
A A
0 T (*,0)
[B ] C [B; D] C
*
D
Tree growth need not involve addition of tree nodes. In a T -structure T
some `underspecied' tree relation may be further specied. For instance, a
pair n; m 2 T such that n D m may be mapped by f to a pair such that
f (n) f (m) and later, skipping # , by some g, to a fully specied relation
g(f (n)) 0 g(f (m)):
A A A
D T (D,*) T (D,*,0)
B C B C B C
Compare this sequence to the following growth stages:
A A A
0 T (*,0) T (D,*,0)
B C B C B C
Here, no actual growth takes place, although each new T -structure is a proper
extension of the previous one. In contrast with the previous growth sequence,
only conclusions of bookkeeping activity are added. Once we have T n 0 T m,
then it is a matter of bookkeeping to add the pair hT n; T mi to the relations,
# ; and D . So, independent of annotations on tree nodes, there is a notion
of growth or development towards complete structures.
Remark 3 (Partial Trees in Normal Form) The partial trees in P P T will,
in our model, all occur in normal form. This form has to do with the under-
specied tree relations #; and D . Consider the situation where we have a
node with a `dangling' daughter and a second node with a daughter.
Declarative Structure 277
(X ) A (Y ) A
U #
B C
Then there are three normal extensions embedding (X ) and (Y ):
(Z1) [A; B ] (Z2) A (Z3) A
# (#,U) #
C or [B; C ] or U C
U
B
That is, either the dangling node B unies with one of the nodes A or C , or it
comes to dangle below the bottom node.
Normal form constraint I Suppose we have T n1 j T n2 for j 2 f; Dg.
If tree T is in normal form and we have T n1 i T n0 for i 2 f0; 1; L; #g, then
there is some T n3 such that T n1 i T n3, and one of the following three is the
case: T n2 = T n1, T n2 = T n3, or T n3 j T n2.
When a structure is in normal form, then an underspecied tree relation is
pushed as low is possible. So, a structure of the form
(Z ) A
* #
C B
is not in normal form. But notice that Z is an -lower bound of the normal
forms Z1; Z2 and Z3. We will consider P P T only to contain trees in normal
form. By proposition 1 clause 2 below, this does not exclude any completions
to full binary trees, for the three possibilities exhaust the possible embeddings
of (X ) in an eventual complete binary tree.
With respect to the relation #, there is the following constraint. Given the
structures
(Y ) A (Z ) A
# and 0 1
B D E
there are only two possible embeddings in normal form
A A
(#,0) 1 or 0 (#,1)
[B; D] E D [B; E ]
Normal form constraint II If a tree node T n1 has a function daughter and
an argument daughter, then the relation T n1 # T n2 entails either T n1 0 T n2
or T n1 1 T n2.
278 The Formal Framework
That is, in case there are function and argument daughters, then # is not an
underspecied tree relation, but an underspecied description of a fully specied
relation. So, the structure
A
# 0 1
B D E
is not in normal form. Notice that every complete binary tree is in normal form
{ that is, satises both constraints.
A pointed partial tree T n is a pointed binary tree if the set fn0 2 P P T j
T n T n g (i.e. the sub-tree of T n with top node n) is a binary tree. These
0
pointed binary trees, elements of P BT , represent a complete sub-term structure
of some logical form under construction. These objects are the goals of our
parsing construction. Thus, the growth of structure of logical forms is handled
in a structure described by the triple
hPPT ; T ; P BT i:
where PPT is a proper subset of the set of all pointed partial trees, and the set
P BT consists of T -maximal elements; that is,
1. for every T n 2 P P T , there is some T 0n0 2 P BT such that T n T T 0 n0,
2. if T n 2 P BT and T n T T 0n0, then T 0 n0 T T n.
Proposition 1 Every structure T n in P P T can be identied with a nite set
of Nor(T n) (of normal forms), the elements of which satisfy constraints I and
II, such that
1. T n is a -lower bound of Nor(T n),
2. if T n 00 T000n0 and0 T 0 n0 2 P BT , then there is some T 00n00 2 Nor(T n) such
that T n T n ,
3. if T 0n0 satises constraints I and II and T n T 0n0 , then there is T 00n00 2
Nor(T n) such that T n T n .
00 00 0 0
For instance, Nor(Z ) = fZ1; Z3g. (Notice that Z3 Z2, but Z2 6 Z3.)
9.2.1.2 F e-structures
The bare pointed partial tree structures are going to be decorated with informa-
tion in the form of labels and formulae. We will dene these decorations in the
general format of feature structures.
A feature structure is a labelled decorated directed graph. The elements of
feature structures will be called points, and to say that feature structures are
labelled directed graphs simply means that one can envisage the points being
Declarative Structure 279
linked by arrows bearing a label. The most important constraint on feature
structures concerns these labelled arrows: there is no point in any feature struc-
ture from which two distinct arrows bearing the same label emerge. In eect,
labelled arrows are representations of partial functions; these partial functions
are usually called features. As to the decorations, we envisage the points of fea-
ture structures being adorned with pieces of information of linguistic interest.
Let's make this precise. We assume signature hF ; D; !i where F and D
are non-empty nite or denumerably innite sets, the set of features and the
set of decorations respectively, and ! : F 7! P (F [ D) is a function satisfying
8f 2 F : !(f ) 6= ;. The function ! assigns the sub-features or values a feature
can take.
Typical elements of F , as used in this book, might be Ty, Fo, Tn, standing
for `type', `formula' and `tree node' respectively, and more linguistic features
like number, person, and tense; while typical elements of D might be e ! t
(2 !(T y)), Walk(Man) (2 !(F o)), 010 (2 !(T n)), and the more usual singular,
plural, 1st, 2nd, 3rd and present. A computational linguist would probably
depict a feature structure as follows:
2 3
Type e!t
4 Formula xW alk(x) 5
Tree Node 01
Here the left column displays elements from F , and the right column elements
from D. Such two-dimensional objects are called Attribute Value Matrices
(AVMs).
As in the case of T -structures, our target decorations, complete feature struc-
tures, declarative units, are special kinds of partial feature structures which are,
in their turn, special kinds of F e-structures.
Denition 5 (F e-structures) An F e-structure of signature hF ; D; !i is a triple
F of the form hU; fRf gf 2F ; fQdgd2D i, where U is a non-empty set; for all f 2 F ,
Rf is a binary relation on U that is a partial function; and for each d 2 D, Qd
is a unary relation on U .
The above AVM can be displayed as an F e-structure of the following shape
RT y RTn RFo
e!t 01 xW alkx
A function g mapping F e-structure F to F0 is a F e-morphism if for all u; u0 2U
if uRf u0, then g(u)Rf0 g(u0), and if u 2 Qd, then g(u) 2 Q0d. We say that F e-
structure F0 extends structure F, notation F F F0 , if there is an F e-morphism
mapping F into F. A function g is an F e-isomorphism from structure F to F0
if g is a bijection and both g and g 1 are F e-morphisms.
For instance,
RT y RT n F RT y RTn RFo F RTy RTn RFo
U 01 e!t 01 e!t 01 xW alkx
280 The Formal Framework
Formula U
Lab a uF Formula John
= Formula U
Just as the pointed partial tree structures in P P T can reach maximal elements
in P BT , so the partial feature structures can be developed into declarative units
DU F S . The elements of DU are feature structures that are in some sense
`complete'. For instance, we have the stipulation that no value of the Formula
feature can occur without a value on the Type feature, so [F ormula John] is an
element of F S but not of DU . The declarative units have in our set-up the role
that sort-resolved and totally well-typed AVMs play in HPSG. Consequently we
can describe the space of partial feature structures by the triple
hP F S; F ; DU i
where P F S is a set of partial feature structures such that hP F S; F i is a meet
semi-lattice and DU P F T is a set satisfying
for every structure F 2 P F S there is F0 2 DU : F F F0 ,
F F F0 for F 2 DU implies F0 F F.
9.2.1.3 FT-structures
In this section we will dene the full structures consisting of pointed partial tree
structures the nodes of which are decorated by F e-structures.
Denition 7 (Feature-decorated T-structures) By a system of (nite, or-
dered, binary) feature-decorated PPTs (fds) (of signature hF ; D; !i) is meant
a tuple hPPT ; (Z; z)i where PPT is a system of pointed partial trees, Z is a
function that assigns to each element T n of P P T a nite, point-generated par-
tial feature structure (of signature hF ; D; !i), and z is a function that assigns
to each node T n of P P T a point z(T n) 2 Z (T n) that generates z(T n).
So at a tree node T n, an F e-structure has the form:9
9 F e-structures in general may have endpoints further than once removed from the gener-
ator. These would represent nested features.
282 The Formal Framework
z (T n)
Z (T n) =
a1 a2 ::: :::
Now, apart from extension of tree relations, formalized by T , we can also con-
sider extensions of node decorations. This notion will simply lift the the relation
F , dened for F e-structures, to the nodes decorated by these structures.
Denition 8 (F e-extensions) A ppt T 0 n0 F e-extends T n, notation T n F
T 0 n0 , if there is a F e-morphism g mapping Z (T n) to Z (T 0 n0 ) such that
g(z (T n)) = z (T 0 n0 )
So T n F T 0n0 implies Z (T n) F Z (T 0 n0). Now a tree node T 0n0 F e-extends
T n if the decoration of T n is updated. The relation F will represent the
ow
of computation on decorations, i.e. feature values. In general, the values of
labels and formula features decorating the various nodes in one tree are not
unrelated but are constrained by co-occurrence principles. This means that
some combinations of (atomic) decorations are simply not represented in the
class of possible decorations, and some combinations of these decorations cannot
co-occur as decorations within the same tree. Thus, updating the value on one
feature dimension may require change of other values. In particular, if the same
value occurs at various nodes in some tree, then updating this value at one
location must mean updating it everywhere. We will formulate the basic notion
of value sharing:
Denition 9 (Value Sharing) Tree nodes T n and T 0 n0 share values on the
features f F , notation T n f T 0 n0, if and only if T = T 0 and for all f 2 f
and b 2 !(f ) we have
z (T n)Rf b () z (T 0 n0 )Rf b:
This relation of value sharing between nodes we will want to make invariant
under extensions. That is, if we replace a meta-variable U by a proper value
at some pptm T n, then every occurrence of U at every node in ppt T must be
simultaneously replaced by that value.
Combining pointed partial tree systems and feature-decorated pointed par-
tial trees gives us our target structures.
Denition 10 (Decorated Pointed Partial Trees) A structure of decorated
pointed partial trees (dppt) relative to signature hF ; D; !i is a quintuple
hP P T; fij i 2 I g; (Z; z ); ; F BT i;
where
1. hP P T; fij i 2 I gi is a system of pointed partial trees,
Declarative Structure 283
2. hP P T; (Z; z)i is a feature-decorated T -structure,
3. =T \ F ,
4. F BT = fT n 2 P P T j T n 2 P BT; Z (T n) 2 DU g.
Here F BT is the set of Feature-decorated pointed Binary Trees, and the order-
ing satises the following constraints:
a. for all f F , if T n T 0n0 and T n f T m, then there is T 0m0 such that
T m T 0 m0 and T 0 n0 f T 0 m0 ;
b. for all T n 2 P P T there is an T 0n0 2 F BT such that T n T 0n0;
c. if T n 2 F BT and T n T 0 n0, then T 0n0 T n.
So the growth relation incorporates both growth of tree structure and growth
of feature decorations.10
10 There are various forms of interaction possible. For instance, typical rewrite steps in-
volving types would be:
[T ype (e; e ! t)] F [T ype t] and [T ype e] F [T ype (e ! t) ! t].
These rewrite steps are known, for instance, from the Lambek Calculus. The typing at tree
nodes will re
ect their combinatorial properties, and in particular their status as function
or argument. Thus, from our perspective, there is an essential dierence between the two
rewrites above. The tree structures on which we will evaluate our languages will be ordered
binary trees, and this ordering will consist in the presence of a function daughter (1 ) and an
argument daughter (0 ). This distinction must correspond to the types decorating the nodes.
The second rewrite, however, transforms an argument into a function; that is, an update along
F interacts with tree structure, and thus with the relation T .
In the situation,
0 1
T y (e) T y (e ! t)
no update of T y(e) to T y((e ! t) ! t)) is possible; that is,
T y (t)
0 1 6 0 1
T y ((e ! t) ! t) T y (e ! t) T y ((e ! t) ! t) T y (e ! t)
due to the con
ict between the decorations at the nodes and the labelling of the arcs. On the
other hand, suppose we have created daughters only with the immediate dominance relation,
then the update is possible. That is,
# # # #
T y (e) T y (e ! t) T y ((e ! t) ! t) T y (e ! t)
T y (t)
1 0 1 0
T y ((e ! t) ! t) T y (e ! t) T y (e ! t) ! t)) T y (e ! t)
284 The Formal Framework
is undened. And, generally, these will be the atomic formulae we will use.
We will, however, give a more general denition which allows decorations with
arbitrary feature structures.
Denition 11 (Atomic DUF Formulae) Let F = hF ; D; !i be a (feature)
signature, let Lai be a letter called a label or a feature identier which can
apply to values a 2 DLa = D \ !(Lai), and let F o be a Formula identier
which applies to values in DF o. Let MV be a set of meta-variables. Now, the
i
In this denition, the bold variables x; y; z should not be confused with their
non-bold variants. The latter occur in the values of the F o feature, the elements
of DF o, in the eventual logical form.
In the previous chapters we have used the connectives and operators
!; 8x; [#], and [# 1 ]. Here we will introduce them by denition.
9.2.1.5 Truth
The truth denition that follows is divided into three parts. First, we will
dene truth for atomic DUF formulae. This denition will involve a twist to
guarantee persistence of atomic formulae in an environment where these may be
updated. Secondly, we will give a purely classical truth denition for the Boolean
connectives, the quantiers and the tree modalities. Here we will introduce
connectives and operators that do not occur in chapter 2. The formulae thus
introduced are interpreted completely on the individual partial trees without
reference to their possible extensions. Thus they will generally not persist
along the relation of extension or growth.
To represent the persistent formulae of chapters 2 and 3, we will, thirdly,
introduce a modality `2', interpreted along , such that 2 represents a formula
persistent under growth. By means of this modality, we will be able to define
the persistent connectives, quantiers and modalities of chapter 2.
Denition 14 (Truth for Atomic DUF Formulae) For
2 AT OM = AT OMF [ f>; ?; }g;
the valuation VP 2 P (AT OM )P P T is dened by:
Declarative Structure 287
if 2 AT OMF , then 2 Vp (T n) i there is a T 0 n0 : T 0 n0 F T n
& Z (T 0n0); z(T 0 n0) j=F .
where j=F (dened below) is the truth relation for feature structures. For
any feature structure F (i.e. hU; fRf gf 2F ; fQdgd2D[MV ; i) and any point
u 2 U:
F; u j=F d i u 2 Qd ;
F; u j=F y i :9d : u 2 Qd ;
F; u j=F hf i i 9u0 (uRf u0 and F; u0 j=F ):
if = >, then 2 Vp (T n).
if = ?, then 62 Vp (T n)
} 2 Vp (T n) i for all n0 6= n 2 T : } 62 V (T n0 )
Remark 5 The formulation of truth for atomic propositions has this form in
order to guarantee persistence of the relation `j=' along in a context where
these propositions may be updated. For instance, if
Z (T n) = Formula John
::: :::
then it may still be that hPPT ; T ni j= F o(U), because
Formula U Formula John
::: ::: F ::: :::
The left AVM may decorate some T 0n0 such that T 0n0 F T n. So, if the node
T 0 n0 has been introduced such that it satises F o(U), then this formula persists
along even if we substitute some value, e.g. John.
The verum symbol > holds at all nodes, the falsum symbol ? holds at no
node. The atomic proposition } represents the pointer and holds at { points
to { exactly one node in every partial tree.
Notice that the pointer } assigns to every partial tree T the unique pointed
partial tree T p, where p is the unique n 2 T such that } 2 V (T n).
There are no surprises in the truth denition for the standard logical constants,
connectives, quantiers and modalities.
Denition 15 (Truth Denition for DUF ) Given a dppt PPT = hP P T; fij
i 2 I g; (Z; z ); ; F BT i, an T n 2 P P T , and # 2 I = f#0 ; #1; #; ; #; L; Dg , then
the relation `j=' between the tuple hPPT ; T ni and a formula 2 DUF is given
by the following clauses:
288 The Formal Framework
hPPT ; T ni j= C () C 2 Vc (T n):
Notice that this denition allows control features to disappear from valuations
along the relation . Moreover, because CF and AT OM are disjoint sets, the
denition of FA(T n) does not include control features.
If we allow non-persistent features, then we must be clear when we may
and when we may not use them. We may not use them to denote aspects of
the tree that are contributions of rules or lexical actions; it would go against
the principle of incrementality to use them for this purpose. On the other
hand, we may use them as conditions guarding actions. Actions may test for
features of the current structure, as there need not be any assumption that these
features are permanent. We will allow classical negation, but only conned to
test conditions for actions.
9.2.2 Goal-directedness
Now we introduce requirement functions, in order to deal with the goal-directedness
of the parsing process, the development of pointed partial tree models towards
fully decorated binary trees.
9.2.2.1 FTR-structures
Denition 17 (Requirement Functions) Given a dppt PPT , a requirement
function V? : P P T 7! P (P DUF ) over PPT is a function assigning nite sets
292 The Formal Framework
Given a specic feature of the logical forms we are interested in, the rst
question is now always: can we devise a system of requirement introduction
such that the fullment of all requirements annotating a given tree corresponds
exactly to a completing of this tree to a term of our typed lambda calculus with
the desired feature? We will give three examples.
Binary trees As a rst example, we will consider a requirement strategy to the
eect that the set F BTV? coincides with the set of binary tree structures. That
is, no underspecied tree relations of the form or # are left in F BTV? that
are not the re
exive transitive closure of the immediate dominance relation
or the union of argument and function daughter relation, respectively. The
idea is to introduce a tree node label, a monadic predicate T n with values in
DT n = fa x; x j a 2 A; x 2 f0; 1; Lg g. That is, a value of this predicate is
a nite sequence of elements of f0; 1; Lg possibly preceded by a constant from
a set A. When trees are constructed in the parsing process, in general it is
not known whether a description that starts o as a top node will remain so
(and thus be the root node of the eventual tree). This is why we introduce a
new node invariably with address a 2 A, satisfying a formula T n(a), where a
is a constant not yet occurring in the construction. In interaction with the tree
modalities various constellations are expressible. So, given the formula T n(a),
expressing the location of a node in the tree under construction, we can x:
T n(a0) $ h"0 iT n(a),
T n(a1) $ h"1 iT n(a),18
and we can x the root node of a tree as follows:
T n(0) $ ["]?:
(Signicantly, we do not internalize the underspecied modalities h#i and h#i
as values of the T n predicate.) The tree node formula T n(0) holds at a node if
it is a top node and remains so throughout all developments. (Note the use of
the falsum { `At every node above the current one ? holds.' As ? is satised
by no node at all (denition 15), this means that there are no nodes above the
current one.)
Now, when we introduce a tree node with an underspecied relation to the
source node, we add the requirement for 9xT n(x) to the node with underspec-
ied address:
[?T y(t)] ) [?T y(t)]
will denote the pptm T with nodes [n1 : : : : : : ] : : : [n : : : : : : ] added and the
pointer now at node n.
k
[n ; T y(e)]
These pptm are not T -morphic, but either can develop into the other, and mov-
ing between them does not change the set of possible (successful) developments
(N1 =LF N2).
19 For the denition of `LF', see remark 7.
Declarative Structure 299
9.2.3.1 Minima and Atoms
A -minimal element T n of PPT M is a one-element domain with empty tree
relations and empty Vp (T n) and V? (T n).
f[n ; ;]gn:
The domain cannot be empty, for, by denition, n 2 T . The minimal element
is the = equivalence class of these elements (there is only one).
The -atoms are the one-element domains T n with empty Vp (T n) and
V? (T n) = fAg for some A 2 DUF (?).
f[n ; A]gn;
where the tree domain T = f[n; A]g. The atoms are single tree nodes with an
atomic requirement. An atom of a special kind will be the Axiom:
Axiom f[n; T y(t)]gn;
where the requirement is for an object of type t, i.e. a proposition. This Axiom
will be the starting point of every parse sequence.
Note that
(T [n ; A]) (T [n A ;]=n);
and not the other way round: requirements lie below facts. The -atoms will
be the starting point of the parsing process; every node of the eventual tree has
come into being as an atom of the form [nT y(t)], [n T y(e)], etc. The typical
atoms are nodes with a requirement for some type.
Note also, for arbitrary A 2 DUF (?),
T [n ?A] T [n A]=n:
All question marks heading a requirement can be got rid of.
9.2.3.2 Joins
The lowest upper bound of T n and T 0n0, denoted by T n t T 0n0 if it exists,
constitutes a -lowest00pptm T 00 n00 above both pptms, such that there are T -
morphisms f : T 7! T ; f : T 0 7! T 00 satisfying f (n) = f 0(n0 ) = n00 and for all
0
m 2 T; m0 2 T 0 such that f (m) = f 0 (m0 ):
Vp (T 00 f (m)) = Vp (T m) tF Vp (T 0 m0 ),
V? (T 00 f (m)) = (V? (T m) [ V? (T 0 m0 ))=FA(T 00 f (m)).
For example:
T [n A A ^ B ] t T [n B A ^ B ]=n = T [n A; B ;]=n:
Notice that N1 = [n A ^ B] and N2 = [nA; B] behave dierently, have dierent
-successors, (N2 [n A B ]; N1 6 [n A B ]), although we have LFV? (N1 ) =
300 The Formal Framework
LFV? (N2 ). For facts annotating tree nodes, the join operation may involve
update of feature values. For requirements, the join operation takes the
requirements of both except for those that are satised by facts.
Because of the demand for the T -morphisms f : T 7! T 00 ; f 0 : T 0 7! T 00 mapping
T n and T n0 to T 00 n00 that they satisfy f (n) = f 0(n0 ) = n00 , if the join exists,
then the images of the two models must share a node.20 This gives us essentially
three kinds of information merge.
1. We can extend a pptm downwards, attaching new nodes to the current
frontier.
This will be used to incorporate subcategorized information (transitive
verbs create a node for a two-place relation and for an object node below
a node requiring a predicate node), and also to add a determiner-requiring
node and a nominal-requiring node below a node requiring type e in case
the natural language string supplies a determiner as decoration for a node
where a formula of type e is required.
2. We may add nodes above the current top node, extending the pptm
upwards.
This will be used in case a clause has been built which turns out to be
a relative or complement clause (as in Japanese). And it will be used in
genitive constructions where a noun phrase sequence turns out to be a
sub-sequence of a determiner (as in English).
3. We may merge nodes occurring in one and the same pptm.
This will be used for incorporation of information from the lexicon, inessen-
tially, and for the resolution of long-distance dependencies essentially.
The incorporation of information from the lexicon can be handled by using
the join operation. If we want to incorporate atomic as a fact and as a
requirement at node T n, then
T n t f[n ]gn
does the trick. Notice that if the requirement already holds at T n, then the
result is the same as
T n t f[n ;]gn:
For instance, the accusative case on some noun phrase contributes a requirement
?h"iT y(e ! t) to the type e node. If the node in question is indeed a daughter
of a node of type e ! t (as will happen in parsing him in John hits him) then
the requirement evaporates. On the other hand, if this node does not occur
20 Both the denitions for join and the meet given here can be extended to partial trees by
the stipulation: if # 2 fu; tg, then T #T 0 () T p#T 0 p0 , where p; p0 are the nodes in T
and T 0 respectively, such that T p j= } and T 0 p0 j= }.
Declarative Structure 301
in such a daughter position (as in parsing Him in Him, John hit), then it will
remain as a requirement until the node has found its position.
Moreover, if F F F0 , i.e. F e-structure F0 is the result of an update of
F e-structure F, then
T [n F ] t f[n F0 ;]gn = T 0 [n F0 ];
where T 0 is the pptm T with every occurrence of F replaced by F0. By denition
10.4a, if more than one of its nodes is decorated with F, then T has no -
successors where only the occurrence of F at T n has been replaced. This kind
of update will take place when substituting, for instance, John for the variable
U in F o(U): Note that
T n t T m;
if it exists, identies the nodes n and m in one and the same tree T . This
cannot be if both n and m are related by fully specied relations (i.e. by
i ; i 2 f0; 1; Lg) to the top node: the position of at least one of them must be
underspecied. In the case of domination this can (and will) be used to unify
nodes. The pptm
T [n ] t T [n 0 0 ];
if it exists, xes an underspecied node in T . Notice that the displayed kind
of unication is typical for T in normal form. If T n and T 0 n0 are in normal
form (remark 3), then their lowest upper bound, if it exists, need not itself be
in normal form. On the other hand, the set Nor(T n t T 0 n0) can represent this
lower upper bound. A normal lowest upper bound of T n and T 0 n0 { that is, an
element in Nor(T n t T n ) { need not be unique.
0 0
9.2.3.3 Meets
The highest lower bound of T n and T 0 n0 , denoted by T n uT 0 n0 , is the -highest
pptm T n such that there are T -morphisms f : T 00 7! T and f 0 : T 00 7! T 0
00 00
such that f (n00) = n; f 0(n00) = n0 and:
Vp (T 00 n00 ) = Vp (T n) uF V (T 0 n0 )
V? (T 00 n00 ) = (V? (T n) \ V? (T 0 n0 ))
[ (V? (T n) \ FA(T 0 n0 )) [ (V? (T 0 n0 ) \ FA(T n))
So the meet of two tree nodes is annotated by the F e-meet of the F e-structures
of the constituent nodes, and has the requirements that appear on both, plus
the requirements of one node that are satised by the other. For instance,
T [n ; ; ;
] u T [n ; ;
]=n = T [n ; ;
]:
Highest lower bounds can be used to deal with varieties of ellipsis and with
structural parallelism in general. In the text John hates Bill; and so does Sue
302 The Formal Framework
it is the lowest upper bound of the representations of John hates Bill and Sue
hates Bill which can be used as the background against which the string so does
Sue is interpreted.
We can dene the notion of a similarity measure to capture structural par-
allelism.
A similarity measure of structure between elements of F BT can be dened by
simT n (T 0 n0 ; T 00 n00 ) meaning that pptm T 0 n0 is more similar to T n than T 00 n00 .
The denition is as follows:
simT n (T 0 n0 ; T 00 n00 ) () T n u T 00 n00 T n u T 0 n0 :
As should be, no element is more similar to T n than T n itself. According to
this measure, for instance, the structure projected by Sue hates Bill is more
similar to that projected by John hates Sue than the structure projected by Sue
likes Bill, and this again is more similar to the rst tree than the one induced
by John walks. A structural similarity class with respect to T n
[T 0 n0]T n = fT 00n00 2 F BT j simT n (T 0n0 ; T 00n00 ) & simT n(T 00n00 ; T 0n0)g
gives a class of decorated binary trees which are structurally similar with respect
to a given target structure. So Jane hates Bill is in the same similarity class
with respect to John hates Bill as Sue hates Bill is.
Another use of the meet operation is the following. Consider the development
sequence
A A A
0 1 T (0,*) 1 T (0,*,D) 1
B C B C B C
Here, it is clear that the addition of information is a matter of mere bookkeep-
ing. It represents a taking into account, making explicit, of the information
present in the previous structure. The structure in which all implicit informa-
tion has been made explicit, can be dened using the u operation. Let the BK
(BookKeeping) function be given by
BK (T n) = ufT 0 n0 2 F BT j T n T 0 n0 g;
where Vp (uD) = uF fVp (T n) j T n 2 Dg, and V?(uD) = f 2 DUF (?) j 9T n 2
D : 2 V? (T n) & 8T n 2 D; 2 V? (T n) [ FA(T n)g. The function BK maps
a structure to the common meet of all developments in which all requirements
have been satised.
9.2.4 Tree Descriptions
Our (theoretical) parser will not work on trees themselves as representations of
partial terms of some logical language, but on descriptions of partial trees.
These descriptions consist of nite sets of DUF (?) formulae. Tree descriptions
will be the computational approximations of partial trees. The approximations
Declarative Structure 303
of tree nodes are nite, splitting, P R-consistent sets of DUF (?) formulae. Here
P R is a set of principles incorporating the Logic of Finite Trees (LOFT) as
well as axiomatizations of the individual predicates or features. For instance,
by standard logical consistency we have no description containing both and
:; by LOFT principles no description contains h#0 i and h#0 i: (`there is
only one argument daughter'); and by axiomatization of the label predicates,
no consistent description of a node contains both T y(e) and T y(t): types are
supposed to be mutually exclusive. We will use the notation `P R if formula
can be computed from formula set using the principles in P R and the usual
derivation rules.
Tree nodes represent partial, dynamic, declarative units: a declarative unit
is a labelled formula, which we represent by a nite set of atomic DUF formulae
fLa1(a1 ) : : : Lan (an ); F o()g satisfying some notion of completeness. A partial
declarative unit is simply a subset of some complete one, and a dynamic partial
declarative unit is a partial one with requirements added. That is, it consists of
a nite number of atomic formulae from AT OM formulae and a nite number
of requirements ? 2 DUF (?) (a valuation Vp and a requirement function V?).
The object is to get rid of the requirements from this set by satisfying them.
Descriptions of these partial dynamic declarative units are generalized sets
of formulae which may include conclusions of this basis; that is, ^ 2 d implies
; 2 d, and _ 2 d implies 2 d or 2 d, and ; ! 2 d implies 2 d.
The tree relations between tree nodes will hold between node descriptions
in virtue of tree node formulae which are generated by a monadic tree node
predicate T n (having values in DT n = fa n j a 2 A [ feg; n 2 f0; 1; Lgg for
some non-empty set A), and by closure under Boolean and modal operations.
Denition 22 (Node Identier Language) For index set I = f#0; #1; #;
#; "0; "1; " g, the language ADR is the smallest subset of DUF (?) satisfying
1. If a 2 DT n [ MV then T n(a) 2 ADR,
2. if 2 ADR, then hii in ADR for i 2 I or i = L.
By means of the node identier language we can now formulate the notions of
node descriptions and tree descriptions.
Denition 23 (Node and Tree Descriptions) A node description d is a -
nite P R-consistent set of DUF (?) formulae satisfying:
1. 9a 2 DT n : T n(a) 2 d,
2. s 2 d ) 62 d,
3. ^ 2 d ) ; 2 d,
4. _ 2 d ) 2 d or 2 d,
5. ; ! 2 d ) 2 d,
304 The Formal Framework
These axioms guarantee that a node has (essentially) only one argument daugh-
ter and function daughter. So we have T D; A `P R h#0i(X ^ :X ), from which
LOFT allows us to conclude T D; A `P R ? in contradiction with the consistency
constraint. So T D is not a tree description.
Notice that the tree relations between two node descriptions are completely
determined by the presence or absence of specic formulae in these descrip-
tions themselves: this tree relation is determined independent of larger tree
descriptions of which these two node descriptions are elements. It may turn
out, however, that the tree description of which node description ND is an ele-
ment determines the position of ND to be more specic than is witnessed inside
ND itself. Consider the following tree description
A = fT n(0) : : : g A
B = fh"iT n(0); X g # 0
C = fh"0 iT n(0); :X g B C
We have A # B, and A 0 C according to denition 23. But there is more
information contained in this tree than is witnessed inside the node descriptions.
By the LOFT principle for # 2 f#; "g,
h#i $ (h#0 i _ h#1 i);
we have that the daughter (mother) relation is the union of the argument
daughter (mother) and the function daughter (mother) relation. We know that
A 0 B or A 1 B ; and, by the functionality axioms, we cannot identify B and
C , thus we must have A 1 B .
T D; fh"iT n(0); X g `P R h"1 iT n(0):
This relationship between node description and tree description will be an im-
portant economy principle, since there will characteristically be information
derivable from the tree description which is not represented in the node descrip-
tions. But this is not all. Consider, nally, the tree description T D consisting
of the nodes A; B; C and D:
A = fT n(0); : : : ; T y(t)g A
B = fh" iT n(0); T y(e); X g * 1
C = fh"iT n(0); [#]?; T y(e); :X g 0
D = fh"1 iT n(0); T y(e ! t)g B C D
In this description, node B has only an underspecied location. We have
A B , A 0 C , and A 1 D, according to denition 23. Again, there is
more information contained in this tree than is witnessed inside the node de-
scriptions. In fact, it must be the case that A 1 D B. For, identifying the
underspecied node B with A or C gives inconsistent sets, and it must dangle
below either C or D. But C can have no daughters, so only D is left to dominate
B . The LOFT principle
h# i $ ( _ h#ih# i)
Procedural Structure 307
regulates this propagation of what we might call the long-distance modalities.
This now allows us to derive
T D; fh"iT n(0); T y(e); X g `P R h" ih"1 iT n(0):
Adding this conclusion to the node description can be seen as a restructuring
of a tree with unxed tree relations.
A
C D
*
B
where we rst make a node at location # removed from the pointer, then we
go there and put there the formula 2 DUF (?) [ CF . Often we will also return
to our point of departure,
make(#); go(#); put(); go(# 1 ):
and
gofirst# (X ) = hfX g; 1; go(#)i ; hfX g; 1; ABORTi:
The action gofirst" (gofirst#(X )) moves to the rst node above (a rst node
below) the starting point where X holds and stays there. So gofirst"(?T y(t))
moves to the top node of the local, minimal clause which the starting point is
contained in.
WHILE X DO = hfX g; ; 1i; hfX g; ABORT; 1i:
This conditional action executes action while X holds at the node of evalua-
tion. So WHILE DO go(") moves to the rst node above the starting point not
in the domain identied by .
If we want to undertake some action at some point not yet known and
return to the current position, we can use control features. Recall that these
features are not persistent, and this is what we want in this case. We only need
a marker to identify the current position; we do not want this marker to be
present in the eventual tree. The idea is to put a control feature C 2 CF at a
node as a requirement ?C (dened as in remark 10). Now we use, for instance,
312 The Formal Framework
Fresh names A further action we will describe will deal with the introduction
of a fresh logical name, either a constant or some variable, to be used in some
formula . The notion of freshness is here dened with respect to the linked
tree structure under development. The action freshput(a; ) generates a xed
nite number of instances of the term a and uses the rst one that is fresh in
. So, to actually contribute a fresh variable to some tree, this number has to
exceed the number of a occurrences in that tree.
freshput(a; )=
IF fhD 1 ihDiF o(a1 )g
THEN IF fhD 1 ihDiF o(a2 )g
THEN : : :
: : : IF fhD 1 ihDiF o(an )g
THEN AB
ELSE put(F o([an =a]))
:::
ELSE put(F o([a2 =a]))
ELSE put(F o([a1 =a]))
This macro is used in the actions projected by a proper name. For instance, the
name John will have the lexical entry
IF ?T y(e)
THEN put(T y(e)); freshput(John; F o(John))
ELSE ABORT
So distinct occurrences of the name John in one sentence will give rise to distinct
logical constants annotating tree nodes. Dierent tree nodes may end up with
instances of the same logical name, but only through a process of copying.
9.3.2 Natural Languages
A natural language will now be represented by an action set L A(ACT ) which
is closed under sequential composition, iteration, non-deterministic choice and
conditional choice. We set L= ( \f P P T M P P T M j 2 Lg). This
is again a partial order: re
exivity follows because the stay put action, `1',
is present. Transitivity follows by closure under sequential composition, and
asymmetry is inherited from . The relation L represents the transitions
within the whole of PPT M that are licensed by the language L.
Every parsing task starts from the Axiom which is a one-node tree, without
tree relations, an empty valuation function, but with a requirement for an object
Procedural Structure 313
of type t:
f[n ; T y(t)]gn
So if a parsing task is fullled, then it must have constructed an object of type t,
a proposition, at this node. By parsing a grammatical natural language string,
the Axiom is connected to an element of F BT by a nite sequence of trees. Each
tree T ni+1 in this sequence is a development of the previous tree T ni , licensed
by the language; i.e. T ni L T ni+1 . So the construction of a full-blown binary
tree (2 F BT ) can be seen as tracing a path of the kind
Axiom = T n0 L T n1 L : : : L T nk 2 F BT:
The set L of actions, is divided into three parts:
1. A set L of L-words. For for every w 2 L the set Lex(w) A(ACT ) is a
nite set of incremental actions.
So a word may have a nite number of (complex) actions associated with
it.
2. A nite set C A(ACT ) of computational actions. These actions we will
generally call rules.
3. A set P A(ACT ) of pragmatic actions.
An action out of each of these sets determines a licensed transition. The actions
in L are generally not dened on, not triggered by, every dppt T n. That is, an
action 2 L determines a subset of P P T M , its domain given by dom() =
fT n 2 P P T M j 9T 0 n0 : (T n; T 0 n0 )g. So every action in L determines a
trigger dom() P P T M . And we will stipulate that a language has only a
nite number of triggers. That is, we stipulate that the set
fdom() P P T j 2 L [ C [ P g
is nite. So the actions associated with L split up the set of decorated partial
trees into a nite number of categories. A language can only be sensitive to a
nite spread of dierences between the decorated partial trees. Note that there
is no limit on the variety in the products of these actions. We may have an
innity of actions, giving an innity of dierent structures, but these actions
share a nite number of domains (or triggers).
The procedural semantics of the parsing process of an L string over PPT M
is now given in terms of LCP congurations. We display an LCP conguration
as a pair
(T n; w)
consisting of a pptm T n and a string w of lexical items; i.e. w 2 L. The set of
all LCP congurations will be denoted by CONF (L; C; P ). In a parse, LCP
314 The Formal Framework
congurations are rewritten to new congurations by lexical actions, computa-
tional rules and pragmatic actions. The basic rewrite relation of our parsing
model is the binary relation,
(T n; w) )LCP (T 0n0; w0 );
where )LCP is dened to be )L [ )C [ )P , and the components of this
union are given by
(T n; w) )L (T 0n0; w0 );
if w = w1 w0 for w1 2 L, 2 Lex(w1), and (T n; T 0n0). That is, )L-transitions
consume the rst word of the string and execute one of the actions it projects.
A transition dened by a computational rule or a pragmatic action has the
form
(T n; w) )x (T 0n0 ; w0)
if x = C0 (0 x = P ), w = w0 , and there is some rule 2 C ( 2 P ) such that
(T n; T n ).
Here no word is consumed; the action undertaken is determined by the set
C of rules or the set P of pragmatic actions.
We may restrict the sets L; C and P to subsets L0; C 0 ; P 0 and consider the
rewrite relation )L0C0P 0 (a subset of )LCP ) with the obvious restricted inter-
pretation. We will drop the empty set from the notation; so, for instance, )C
will be short for );C;. As usual, the relation )LCP will consist of the re
exive
and transitive closure of )LCP .
If language L is determined by L; C and P , then we can x our relation
L =LCP of licensed development or growth by
T n L T 0 n0 () T n T 0 n0 & 9w; w0 2 L : (T n; w) )LCP (T 0 n0 ; w0 ):
(Because pointer movements are included as possible actions, we have to de-
mand that T n T 0n0 holds in the above denition). Analogously, we can
dene relations CP , C , etc. Now we can restrict the general notion of de-
velopment of tree structure as formalized in the relation between partial tree
structures to those developments that are licensed by the lexical actions L and
the computational rules C (see section 9.3.2). That is, we can evaluate logical
constants along developments licensed by a particular natural language. For
instance, an implication
hPPT M; (T n; w)i j= !LCP
will now be evaluated by
for all (T 0 n0 ; w0 ) such that (T n; w) LCP (T 0 n0 ; w0 ), if hPPT M; (T 0 n0 ; w0 )i
j= then hPPT M; (T 0 n0 ; w0 )i j= .
Procedural Structure 315
Moreover, we can restrict this implication to subsets of LCP ; for instance,
CP considers only developments without lexical contributions { that is, if
(T n; w) CP (T 0n0 ; w0), then w = w0 . This is the relation along which the
implication !CP is evaluated; we consider invariances over developments that
are possible without referring to the next word. Of interest is the restricted
negation :C
hPPT M; (T n; w)i j= :C ;
which states that (T n; w) has no successors that can be reached by computation
alone and where holds. On the other hand, lexical actions may still make
such a successor available.
It is now straightforward to dene a concept of well-formedness over the
sequence of transitions which a string of words must dene if it is to successfully
lead to a logical form as one possible interpretation assigned to those words.
The sequence of words must lead from the Axiom to the nal state using each
action dened by the words in the order given, computational rules or pragmatic
actions, so that a well-formed logical form results with no requirements on any
tree-node left outstanding and no words left unprocessed.
(Axiom; w) = (T n0; w) )LCP (T n1; w0 ) )LCP : : : )LCP (T nn; e) 2 F BT:
Recall that e denotes the empty string.
Now, the set of grammatical strings, given C , L and P , can be dened as the
strings that, starting from the Axiom, will reach an element of F BT exactly at
the last word.
Denition 29 (Grammatical L-strings) For L a natural language, the set
ST RU (L) of all L-structures is given by
ST RU (L) = fT n 2 P P T j 9T 0 n0 2 F BT : Axiom LCP T n LCP T 0 n0 g:
Language L is expressively adequate (EA) if
F BT ST RU (L):
If natural language L is expressively adequate, then it can project any logical
form in our representation language. Given a natural language L, the set of
L-grammatical strings, L(L; C; P ) is given by either of the two stipulations
1. fw1 : : : wn 2 L j 9T n 2 F BT : (Axiom; w1 : : : wn) )LCP (T n; e)g,
2. fw1 : : : wn 2 L j 9i 2 Lex(wi ); 1; : : : ; n+1 2 C [ P j
j hPPT M; Axiomi j= h1 ; 1 ; 2 ; : : : ; n ; n ; n+1 i(["]?^?;)g:
The representations projected by a string s 2 L(L; C; P ) are given by
Repr(s) = fT n 2 F BT j (Axiom; s) )LCP (T n; e)g:
Depending on the level of underspecication of s { that is, the use of prag-
matic actions in P { the set Repr(s) may be a singleton (All men hate chocolate),
a nite set (Every dog hates a man), or an innite one (He hates him).
316 The Formal Framework
Computational rules
By denition, whenever we have (T n; w) )C (T 0n0; w), then we have 8w 2 L :
(T n; w) )C (T 0n0; w), because the string is not involved in computational rules.
So the notion T n )C T 0n0 makes sense. Computational rules will be based on
principles of the form
! hi>:
Such a formula claims the existence of an action transition, given the right pre-
condition. Each principle ! hi> is associated with a relation !hi> 2 C ,
given by:
!hi> = fhT n; T 0 n0 i j hPPT M; T ni j= & (T n; T 0 n0 )g:
The progress from Axiom to F BT is non-deterministic: at every state of the
parse the word currently under consideration can generally be assigned more
than one structural role in the tree by principles in C [ P . For instance, a noun
phrase sequence initiating a natural language string may end up as subject
John, admires, Mary; it may be a fronted object, Mary, John, admires; or it
may be a topic constituent, (As for) Mary, John, admires, her. The set C must
contain actions to set up these three structures. The transitions determined by
the computational rules, i.e.
(T n; w) )C (T 0n0; w)
are non-deterministic ()C is a multi-valued function). But, as is evident, for
the parsing process to get o the ground at all, for any given tree T n the set C
should contain only a nite number of (alternative) actions which have T n in
their domain: only a nite number of developments can be computed. Moreover,
we should not be able to apply rules indenitely: after a certain number of
applications, we must have reached a tree that is no longer in the domain of any
rule in C . Thus, for a given T n, not only is fT 0n0 j 9w : (T n; w) )C (T 0 n0; w)g
nite, but also fT 0n0 j 9w : (T n; w) )C (T 0n0 ; w)g.
We will represent rules of the form ! hi> by
Tn
T 0 n0
where hPPT M; T ni j= and (T n; T 0 n0).
Pointer movements
Computational pointer movements are pointer movements not induced by lexical
information. We allow the pointer to leave a node by a computational rule only
when that node has no outstanding requirements. If this is the case, then a
computational pointer movement consists of a (complex) action composed of
go(#) actions, the regular operations and conditionalization:
?; ! hi>:
Axioms 317
We know that hPPT M; T ni j=?; implies that hPPT M; T n0i j=?; for every
T n i T n0 where i 2 f0; 1; #; g; thus movement downward from a node without
requirements is always allowed.
1. ?; ! hgo(#1 ); : : : ; go(#n )i>
moves the pointer from a node with no requirements along a path in the
tree specied by #1 : : : #n (note that along this path no test is made for
absence of requirements). Of course, this action presupposes the presence
of this path.
2. ?; ! hhfh"i>g; go(h"i); 1ii>
moves the pointer one up, if there is such a node.22
9.4 Axioms
In this section we present the axiomatization of the logic of nite ordered binary
trees, of partial trees, of requirements of action modalities and we will present
a selection of computational rules and uodate actions.
9.4.1 Finite Binary Trees
As the basis, we take any suitable axiomatization of the intuitionistic predicate
calculus, together with all instances of the schemas B1{B10, E1{E4, D1{D11
and F1{F3 below under the scope of the operator [LF]. As rules of inference we
take Modus Ponens (if and ! are provable, then so is ) and the rule of
necessitation in [#0], [#1], ["], [#], [#], ["], [LF] and G; that is, if is provable,
then so are [#0], [#1], ["], [#], [#], ["], [LF] and G. Formal proofs
are nite sequences of ws built using the axioms and rules of inference in the
usual way. If a w is provable, then we write ` and say that is a theorem.
With these generalities to hand, let us examine the details. The axioms
split naturally into four groups. First of all, there are the axioms for the basic
operators.
All the axioms of the groups B, D, E and F are under the scope of the operator
[LF].
B1 [#0 ]( ! ) ! ([#0 ] ! [#0 ] )
B2 [#1 ]( ! ) ! ([#1 ] ! [#1 ] )
B3 ["]( ! ) ! (["] ! ["] )
B4 h#0 i ! [#0]
B5 h#1 i ! [#1]
22 Notice that the righthand side of the above rule contains a conditional action h ; ; i
within an existential modality h i.
318 The Formal Framework
B6 h"i ! ["]
B7 ! [#0]h"i
B8 ! [#1]h"i
B9 ! ["]h#i
B10 h#1 i> $ h#0 i>
Most of this is familiar from tense/modal logics. B1{B3 are universally valid
modal principles, while B4{B6 re
ect the partial functional (or `deterministic')
nature of the 0, 1 relations. B7 and B8 are familiar from tense logic: they
record the fact that both the converse of 0 and the converse of 1 are contained
in the converse of . B9 is closely related and says that the converse of is
contained in 0 [ 1 (to see this, recall that, on full binary trees (in F BT )
h#i is shorthand for h#0 i _ h#1 i). Finally, B10 takes account of the fact that
function and argument daughters come into being together.
The next group of axioms deals with the transitive closure operators and
their interactions with the basic operators.
E1 [# ] $ ( ^ [#][# ])
E2 [" ] $ ( ^ ["][" ])
E3 [# ]( ! [#]) ! ( ! [# ])
E4 [" ]( ! ["]) ! ( ! [" ])
These are familiar from temporal logic and Propositional Dynamic Logic (see
Goldblatt 1992). They re
ect the fact that is the re
exive transitive closure
of .
The intended meaning of the dened symbols was discussed in the previous
sections; the next group of axioms pins these down.
D1 T n(0) $ ["]?
D2 T n(a0) $ h"0 iT n(a)
D3 T n(a1) $ h"1 iT n(a)
D4 T n(aL) $ hL 1 iT n(a)
D5 h#i $ (h#0 i _ h#1 i)
D6 [#0 ] $ :h#0 i:
D7 [#1 ] $ :h#1 i:
D8 ["] $ :h"i:
D9 [#] $ :h#i:
Axioms 319
D10 [#] $ :h# i:
D11 ["] $ :h" i:
Finally we turn to the axioms that give LOFT its
avour: namely, those that
re
ect the fact that our intended models are all nite.
F1 h# i[#]?
F2 h" iT n(0)
F3 ! h#i( ^ [#][# ]:)
F1 and F2 are straightforward: no matter where we are in a nite tree, we are
only a nite number of backwards steps away from the root node (which is
what F2 says) and a nite number of steps away from at least one terminal
node (which is what F1 says). More interesting is F3. Roughly speaking, it says
that if holds at any node in a tree, then this node dominates a node not
dominating any other nodes. It is this axiom that will enable us to maintain
the niteness of the tree.23
9.4.2 Partial Trees
Now we turn to principles that axiomatize the trees outside F BT . We relate
the elements of F BT to the trees in P P T in general by the following principles:
LF1 hLFi>
LF2 ! [LF]
LF3 [LF]( _ :)
LF4 [LF] ^ hLFi ! hLFi( ^ )
Notice that we know already that any of the above axioms can be preceded by
[LF], G and thus by ::. This severely restricts the partial trees in P P T . As
all partial trees can be embedded in some structure where all the LOFT axioms
hold (by LF1), no circular or innite structures can arise. But we have argued
(in remark 3) that not all partial trees that can be embedded in an element of
F BT should be allowed in P P T : we want P P T to contain only trees satisfying
the normal form constraints. Thus, we add the following principles:
U1 (h#0 i> ^ h#1 i>) ! (h#i ! (h#0 i _ h#1 i))
U2 h#i> ! (h# i ! ( _ h#ih# i))
23 For more information about this axiomatization see Blackburn and Meyer-Viol 1994.
320 The Formal Framework
That is:
If both daughters exist, then the modality h#i is not an underspecied tree re-
lation, but an underspecied statement about a fully specied tree.
If a daughter exists, then the modality h#i can be further specied by one level;
that is, it is partly an underspecied statement (as to the daughter) and partly
a statement about an underspecied tree relation (the relation dangling below
the daughter). That is, an underspecied dominance relation attaches as low as
possible in the tree.
We will also use the standard interaction schema for universal and existential
modalities.
U3 ([#] ^ h#i ) ! h#i( ^ )
9.4.3 Requirements
Now we turn to schemas involving requirements. By denition 17 the following
four schemas hold:
R1 !s?
R2a ?; ! (> !?;)
R2b ?; ! [#]?;
R3 ? 7! 2( _?)
We will also want our models to be `saturated' in the following sense: every
persistent that is neither in F A(T n) nor in V0 ? (0 T n) (i.e. is neither a fact
nor a requirement at T n) is falsied by some T n 2 LFV? (T n); that is,
hPPT M; T 0 n0 i j=s :
The (left-to-right direction of the) following principle expresses this:
R4a G $ (_?)
R4b G?;
Now we have some interaction principles:
R5 G ^ F ! F ( ^ )
R6 F ! hLF i
R7 [LF ] ! G
Axioms 321
9.4.4 Actions
A next group of axiom schemas involves the actions in A(ACT ):
A0a h1i $
A0b [ABORT]?
A1a [put()]
A1b ! hput()i
A1c hput()i ! [put()]
A2 [make(#)]h#i>
A3a ! [go(#)]h# 1 i
A3b hgo(#)i ! h#i
A4a _ h# i ! [merg()]
for ; without question marks
A4b ? _ h# i? ! [merg()](_?)
A4c hmerg()i ! ( _ h# i)
A4d hmerg()i ! [merg()]
The next group of axioms are the standard PDL ones xing the regular opera-
tions. Our axiom for the conditional test, however, is new.
A5 [1 ; 2 ] ! [1 ][2 ]
A6 [1 + 2 ] ! [1 ] ^ [2 ]
A7 [ ] ! ^ [][ ]
A8 [ ]( ! []) ! ( ! [ ])
A9 [h; 1 ; 2 i] ! 8x( ! [1 ]) _ (s9x ^ [2 ])
where x contains all free variables in .
We also have the functionality axioms:
hmake(#)i ! [make(#)]
hgo(#)i ! [go(#)]
if # 2 fh#0i; h#1i; h"ig.
322 The Formal Framework
Fodor, J. A., and J. Lepore. 1992. Holism: A Shopper's Guide. Oxford: Black-
well.
Fodor, J. D., and I. Sag. 1982. Referential and quanticational indenites.
Linguistics and Philosophy 5, 355{398.
Gabbay, D. 1996. Labelled Deductive Systems. Oxford: Oxford University Press.
Gabbay, D., and I. Hodkinson. 1996. Temporal logic in the context of databases.
In Logic and Reality, ed. J. Copeland, 69{87. Oxford: Oxford University
Press.
Garrod, S., and A. Sanford. 1981. Bridging inferences and the extended domain
of reference. In Attention and Performance IX, ed. J. Long and A. Boddeley,
331{346. Hillsdale: Lawrence Erlbaum.
Gartner, H. M. 1997. Are there V2 relative clauses in German? MS, University
of Potsdam.
Ginzburg, J. 1992. Questions, Queries, and Facts: A Semantics and Pragmatics
for Interrogatives. Doctoral dissertation, University of Edinburgh.
Ginzburg, J. 1995. Resolving questions I. Linguistics and Philosophy 18, 459{
527.
Ginzburg, J., and I. Sag. 1999. English Interrogative Constructions. MS, Stan-
ford University and University of Jerusalem.
Givon, T. 1976. Topic, pronoun and grammatical agreement. In Subject and
Topic, ed. C. Li and S. Thompson, 149{188. New York: Academic Press.
Goldblatt, R. 1992. Logics of Time and Computation. Stanford: CLSI Press.
Gorrell, P. 1995. Syntax and Parsing. Cambridge: Cambridge University Press.
Groenendijk, J., and M. Stokhof. 1989. Type-shifting rules and the semantics of
interrogatives. In Properties, Types and Meaning: Vol.2: Semantic Issues,
ed. G. Chierchia, R. Turner, and B. Partee, 21{68. Dordrecht: Kluwer.
Groenendijk, J., and M. Stokhof. 1991. Dynamic predicate logic. Linguistics
and Philosophy 14, 39{100.
Groenendijk, J., and M. Stokhof. 1997. Questions. In Handbook of Logic and
Language, ed. J. van Benthem and A. ter Meulen, 1055{1125. Amsterdam:
Elsevier.
Grosu, A. 1996. Missing-P free relative constructions. Linguistic Inquiry 27,
257{293.
Hale, K. 1983. Warlpiri and the grammar of noncongurational languages.
Natural Language and Linguistic Theory 1, 5{47.
Hausser, R. 1989. Computation of Language. Berlin: Springer.
Hausser, R. 1999. Foundations of Computational Linguistics. Berlin: Springer.
Heim, I. 1982. The Syntactic and Semantic Properties of Denite and Indef-
inite Noun Phrases. Doctoral dissertation, University of Massachusetts,
Amherst.
Heim, I. 1990. E-type pronouns and donkey anaphora. Linguistics and Philos-
ophy 13, 137{177.
Hepple, M. 1990. The Grammar and Processing of Order and Dependency: A
Categorial Approach. Doctoral dissertation, University of Edinburgh.
Bibliography 331
Heusinger, K. von. 1997. Denite descriptions and choice functions. In Logic,
Language and Computation, ed. S. Akama, 61{92. Dordrecht: Kluwer.
Heycock, C. 1995. Asymmetries in reconstruction. Linguistic Inquiry 26, 547{
570.
Higginbotham, J. 1987. The autonomy of syntax and semantics. In Knowl-
edge Representations and Language Understanding, ed. J. Gareld, 119{
131. Cambridge, MA: MIT Press.
Higginbotham, J. 1996. The semantics of questions. In The Contemporary
Handbook of Semantic Theory, ed. S. Lappin, 361{385. Oxford: Blackwell.
Higginbotham, J., and R. May. 1981. Questions, quantiers and crossing.
Linguistic Review 1, 41{79.
Hilbert, D., and P. Bernays. 1939. Grundlagen der Mathematik II. Berlin: Julius
Springer.
Hirschbuhler, P. 1981. The ambiguity of iterated multiple questions. Linguistic
Inquiry 135{146.
Hohle, T. 1996. The w-w{construction: appositive or scope indicating? In
Papers on Wh-Scope Marking, ed. U. Lutz and T. Muller, 37{58. Stuttgart:
University of Stuttgart.
Hoji, H. 1986. Scope interpretation in Japanese and its theoretical implications.
In Proceedings of West Coast Conference in Linguistics, ed. M. Dalrymple
et al., 87{101.
Horvath, J. 1997. The status of `wh-expletives' and the partial wh-movement
construction of Hungarian. Natural Language and Linguistic Theory 15,
509{572.
Hoshi, K. 1995. Structural and Interpretive Aspects of Head-Internal and Head-
External Relative Clauses. Doctoral dissertation, University of Rochester,
New York.
Huang, J. 1982. Logical Relations in Chinese and the Theory of Grammar.
Doctoral dissertation, MIT.
Inoue, H., and J. D. Fodor. 1995. Information-paced parsing of Japanese. In
Japanese Sentence Processing, ed. R. Masuka and N. Nagai, 9{63. New
York: Lawrence Erlbaum.
Jackendo, R. 1992. Languages of the Mind: Essays on Mental Representation.
Cambridge, MA: MIT Press.
Jacobson, P. 1996. The syntax-semantics interface in categorial grammar. In
The Handbook of Contemporary Semantic Theory, ed. S. Lappin, 89{116.
Oxford: Blackwell.
Jiang, Y. 1995. Logical Dependency in Quantication. Doctoral dissertation,
University College London.
Johnson, D., and S. Lappin. 1997. A critique of the minimalist program.
Linguistics and Philosophy 20, 273{333.
Johnson, D., and S. Lappin. 1999. Local Constraints vs Economy. Stanford:
CSLI Press.
332 Bibliography
Joshi, A., L. Levy, and M. Takahasihi. 1975. Tree adjunct grammars. Journal
of Computer and System Sciences 10, 136{163.
Kadmon, N. 1990. On uniqueness. Linguistics and Philosophy 13, 273{324.
Kamp, H. 1984. A theory of meaning and semantic representation. In Truth,
Interpretation, and Information, ed. T. Janssen and M. Stokhof, 1{34. Dor-
drecht: Foris.
Kamp, H. 1996. Discourse Representation Theory and Dynamic Semantics:
representational and nonrepresentational accounts of anaphora. MS, IMS,
Stuttgart.
Kamp, H., and U. Reyle. 1993. From Discourse to Logic. Dordrecht: Kluwer.
Kaplan, R., and A. Zaenen. 1989. Long-distance dependencies, constituent
structure, and functional uncertainty. In Alternative Conceptions of Phrase
Structure, ed. M. Baltin and A. Kroch, 17{42. Chicago: University of
Chicago Press.
Karttunen, L. 1977. Syntax and semantics of questions. Linguistics and Phi-
losophy 1, 3{44.
Kayne, R. 1994. The Antisymmetry of Syntax. Cambridge, MA: MIT Press.
Keenan, E. 1985. Relative clauses. In Language Typology and Syntactic De-
scription: Volume II - Complex Constructions, ed. T. Shopen, 141{170.
Cambridge: Cambridge University Press.
Kempson, R. 1995. Ellipsis in a labelled deduction system. Bulletin of the
Interest Group in Pure and Applied Logics 3, 489{526.
Kempson, R. 1996. Semantics, pragmatics and natural-language interpretation.
In Handbook of Contemporary Semantic Theory, ed. S. Lappin, 561{598.
Oxford: Blackwell.
Kempson, R., and A. Cormack. 1981. Quantication and ambiguity. Linguistics
and Philosophy 4, 253{309.
Kempson, R., and D. Gabbay. 1998. Crossover: a unied view. Journal of
Linguistics 34, 73{124.
Kempson, R., W. Meyer-Viol, and D. Gabbay. 1998. VP Ellipsis: towards a dy-
namic, structural account. In Fragments: Studies in Ellipsis and Gapping,
ed. S. Lappin and S. Benmamoun, 227{290. Oxford: Oxford University
Press.
Kibble, R. 1997. Anaphora Resolution in Dynamic Semantics. Doctoral disser-
tation, University of Edinburgh.
Kim, I. K. 1997. Agreement Phrases in DP. In University College London Work-
ing Papers in Linguistics, ed. P. Backley and J. Harris, 281{302. London:
University College London.
Kiss, K. 1991. Congurationality in Hungarian. Dordrecht: Reidel.
Kuroda, S. Y. 1999. Shubu naizai kankei-setu (head internal relative clauses).
In Kotoba no Kaku to Shuuen, ed. S.Y. Kuroda and M. Nakamura, 17{103.
Tokyo: Kuroshio Shuppan.
Lappin, S., and N. Francez. 1994. E-type pronouns, I-sums, and donkey
anaphora. Linguistics and Philosophy 17, 391{428.
Bibliography 333
Lappin, S., and T. Reinhart. 1988. Presuppositional eects of strong determin-
ers: a processing account. Linguistics 26, 1021{1037.
Lasnik, H., and T. Stowell. 1991. Weakest crossover. Linguistic Inquiry 22,
687{720.
Leisenring, A.C. 1969. Mathematical Logic and Hilbert's Epsilon-Symbol. Lon-
don: MacDonald Technical and Scientic.
Lemmon, E.J. 1965. Beginning Logic. London: Nelson.
Mahajan, A. 1990. The A/A-bar Distinction and Movement Theory. Doctoral
dissertation, MIT.
Maling, J., and A. Zaenen. 1981. A phrase-structure account of Scandinavian
extraction phenomena. In The Nature of Syntactic Representation, ed.
P. Jacobson and G. Pullum, 229{282. Dordrecht: Reidel.
Manzini, R., and S. Savoia. 1998. The syntax of object clitics si in Italian
dialects. MS, University of Florence.
Manzini, R., and K. Wexler. 1987. Binding theory, parameters, and learnability.
Linguistic Inquiry 18, 413{444.
Maracz, L., and P. Muysken. 1988. Congurationality: The Typology of Asym-
metries. Dordrecht: Foris.
Marcus, M. 1980. A Theory of Syntactic Recognition for Natural Language.
Cambridge, MA: MIT Press.
Marcus, M. 1987. Deterministic parsing and description theory. In Linguistic
Theory and Computer Applications, ed. P. Whitelock, M. Wood, H. Somers,
R. Johnson, and P. Bennett, 69{112. New York: Academic Press.
Marcus, M., D. Hindle, and M. Fleck. 1983. D-Theory: Talking about talking
about trees. In Proceedings of 21st Annual Meeting of the Association of
Computational Linguistics, 129{136. Cambridge, MA.
Marten, L. 1999. Syntactic and Semantic Underspecication in the Verb Phrase.
Doctoral dissertation, School of Oriental and African Studies, University of
London.
May, R. 1985. Logical Form. Cambridge, MA: MIT Press.
McCawley, J. 1968. Concerning the base component of a transformational
grammar. Foundations of Language 4, 243{269.
McCloskey, J. 1988. Resumptive pronouns, A' binding and levels of representa-
tion in Irish. In The Syntax of the Modern Celtic Languages, ed. R. Hen-
dricks, 199{248. New York: Academic Press.
McDaniel, D. 1989. Partial movement and multiple wh-movement. Natural
Language and Linguistic Theory 7, 565{604.
Meyer-Viol, W. 1995. Instantial Logic. Doctoral dissertation, University of
Utrecht.
Milward, D. 1994. Dynamic Dependency Grammar. Linguistics and Philosophy
17, 561{606.
Milward, D. 1995. Dynamics and situations. In FRACAS Deliverable D15, ed.
R. Cooper et al., 194{216. Edinburgh: University of Edinburgh.
334 Bibliography
, 42 T , 275
#, 48, 291 h(x); ; 0 i, 310
# , 48, 291 1, 90, 309
[#], 37 j=, 32, 39, 46, 48, 287
[#0], 37, 317 }, 51, 56, 79
[#1], 37, 317 , 39, 274
[#], 37, 318 D , 110, 274
["], 37, 317 # , 274
["], 37, 318 1 , 39, 274
h#1 i, 37, 317 F , 279
h#0 i, 37, 317 L , 110, 274
h#1 i1, 96 0 , 39, 274
h#0 1 i, 37 , 39, 274
h"0 1 i, 37 1 , 96
h#1 1 i, 37 ?, 38, 42, 270, 293
h"1 i, 37 ?;, 38, 293, 320
h# 1 i, 37 !C , 323
h" 1 i, 37 )LCP , 78, 314
h# 1 i, 37 >, 38, 285
t, 299{301 ?, 38, 285