You are on page 1of 8

From: AAAI-80 Proceedings. Copyright 1980, AAAI (www.aaai.org). All rights reserved.

KnowledgeRepresentation
for Syntactic/Semantic
Processing
Robert J. Bobrow
Bolt Beranekand Newman Inc.
50 Moulton St.
Cambridge,Mass. 02238
Bonnie L. Webber
Departmentof ComputerScience
The Moore Schoolof ElelectricalEngineeringD2
Universityof Pennsylvania
Philadelphia,Pa. 19104
Because RUS provides a very clean interface
between syntacticand semantic processing,it has
been possible to experiment with a variety of
knowledge representations in
different
the
implementationsnoted above. The most recent such
implementationuses the KL-ONE formalism c31,
to represent the knowledge needed for
[41,
incrementalprocessing. (This implementationhas
been dubbed PSI-KLONE, for "Earsing and Semantic
InterpretatioGsmL-ONE1l.)
KL-ONE is a uniform
Ebject-centeredrepresentationalscheme based on
the idea of
structured inheritance in a
lattice-structured
taxonomy of generic knowledge.
As we shall discuss iater,- PSI-KLONE takes
advantageof KL-ONE'staxonomiclattice [ill which
ccmbines the propertiesof an inheritancenetwork
with those of-a discrimination
net.

1. Introduction
This paper discusses some theoretical
implications of recasting parsing and semantic
interpretation
as a type of inferenceprocesswhich
we call incrementalhescription refinement.*It
draws upon our recent experience with RUS. a
frameworkfor naturallanguageprocessingdeveloped
at BBN and in use in several different natural
language systems across the country (for details
see [II, C21 and C71>. RUS is a very practical
system that is as efficientas a semanticgrammar
like the SOPHIE parser 161 and as flexible and
extensible
modular syntactic/semantic
processor lik? LUiAR CIOI.
It achieves this
ccmbination of efficiency and flexibility by
cascading Cl21 syntacticand semanticprocessingproducingthe semantic interpretationof an input
utteranceincremental1
y during the parsingprocess,
and using it to guide the operationof the parser.

The next section of this paper describesthe


syntactic/semantic
cascade in general terms, and
then gives a short example of its operationin the
PSI-KLONE implementation. We then define the
concent of an incrementaldescribtionrefinement
(IDR)' process to use as a* paradigm for
usrstanding
the operation of the semantic
componentof-the cascade. This introducesthe last
which discusses the
section of the paw,
requirements for a general frame-like knowledge
representation
if it is to be capableof supporting
such an IDR process.
2. The Syntactic/Semantic
Cascade
Within the RUS framework, the interaction
between the. parser and the semantic interpreter
(the interpreter)takes place incrementallyas the
parser scans the input string from left to right,
one word at a time. The semanticinterpretation
of
each syntacticconstituentis produced in parallel
with the determinationof its syntacticstructure.
Knowledgedevelopedin the course of producingthe
interpretation
is used to controlfurtheraction by
the parser. Parsing supports the processes of
semantic interpretationand discourse inference
(not discussed in this paper) by finding the
constituents of each phrase, determining their
syntacticstS;ucture,
and labellingtheir functional
relationship to the phrase as a whole (the
*

We use an extendednotion of functionalrelation


here that includes surface syntactic relations,
logical syntactic (or shallow case structure)
relations, and relations useful for determining
discoursestructuressuch as primaryfocus.

*The research reported in this paw


was
supportedin part by the .Advanced
ResearchProjects
Agency,and was monitoredby ONR under ContractNo.
NOOO14-77-C-0378.

316

matrix). These labels are proposedpurely on the


basisof syntactimrmation,
but are intendedto
reflect a constituent'sfunctional role in the
matrix, apd not simply its internal syntactic
structure. We will refer to these labels as
functionalor syntacticlabelsfor constituents.
The parser and interpreter engage in a
dialogue consisting of transmissions from the
parser and responses from the interpreter. A
transmissionis a proposal by syntax that some
snecific functional relation holds between a
previouslyparsed and interpretedconstituentand
the matrix phrase whose parsing and interpretation
is in progress. The proposaltakes the form of a
matrix/label/constituent
triple. The interpreter
eitherrejects the proposal or accepts it and
returns a pointerto a KL-ONE data-re
which
represents-itsknowledgeof the resultingphrase.
(This pointeris not analyzedby the parser,but is
rather used in the descriptionof the matrix that
syntax includesin its next proposal(transmission)
to extend the matrix.) The parser is implemented
as an ATN 191, and transmissionsoccur as actions
on the arcs of the ATN grammar. The failureof an
arc because of a semantic rejection of a
transmissionis treatedexactlylike the failureof
an arc becauseof a syntacticmismatch;alternative
arcs on the source state are attempted,and if none
are successful,a back-upoccurs.
2.1. The role of the semanticinterpreterin a
cascadedsystem
The PSI-KLONE interpretermust perform two
relatedtasks:
1.

provide. feedback to the parser by


checking the semantic plausibility of
proposed
syntactic
labels
for
constituentsof a phrase,and

2. build semantic interpretations for


individualphrases
The mechanismfor performingboth these tasks
is based on the idea of mapping between the
(syntactic) functional labels provided by the
parserand a set of extendedcase-frameor semantic
relations (defined by the inters)
that can
hold between a constituentand its matrix phrase.
The mapping of functional labels to semantic
relationsis clearlyone to many. For example,the
logical subject of a clause whose main verb is
"hiV1might be the agent of the act (e.g. "'Iheboy
hit ...I'> or the instrument(e.g. "The brick hit
A semanticrelation (or semanticrole), on
. . .111.
the other hand, must completelyspecify the role
played by the interpretation
of-the constituentin
the interpretation
of the matrix phrase.

The task of the interpreteris to determine


which, if any, semanticrelationcould hold between
a matrix phrase and a parsed and interpreted
constituent,given a functionallabel proposed by
the parser. This task is accomplishedwith the aid
of a-set of pattern-actionrelationmapping rules
(RMRULES)that specifyhow a given funlabel
cmapped
into a semanticrelation. An RMRULE
has a pattern (a matrix/label/constituent
triple)
that specifiesthext=in
which it applies,
in terms of:
0

the syntactic shape of the matrix (e.g.


"It is a transitiveclausewhose main verb
is 'run'."), and the interpretationand
assigned to
other
semantic role
constituents (e.8. 'IThelogical subject
must be a person and be the Agent of the
clause"),

the proposedfunctionallabel, and

the interpretationof the constituentto


be added.

The action of the RMRULE is to map the given


functionallabelonto a semanticrelation.
A proposed syntactic label is semantically
plausible if its proposal triple matches the
pattern triple(s)of some RMRULE(s). KL-ONE is a
good language for describing structured objects
such as phrases built up out of constituents,and
for representingclasses of objects such as the
matrix/label/constituent
triples that satisfy the
constraints given by RMRULE patterns.
In
PSI-KLONE,each RMRULE pattern is representedas a
KL-ONE structure called a Generic Concept (see
section 2.2). These Concep55??5- am
in a
taxonomy that is used as a discriminationnet to
determine the set of patterns which match each
triple. We refer to this as the taxonomy of
syntactic/semantic shapes; note that it is
generally a lattice and not simply a tree
structure.
Associatedwith each semantic relation is a
the
rule (an IRULE) that specifies how
interpretation
of the constituentis to be used in
building the interpretationof the matrix phrase.
When all the constituentsof a matrix have been
assigned appropriate semantic relations, the
interpretation
of a phrase is built up by executing
all of the IRULEs that apply to the phrase. The
separationof RMRULEs from IRULEs allows PSI-KLONE
to take full advantage of the propertiesof the
As
each new
syntactic/semantic cascade.
constituent is proposed by the parser, the
interpreteruses the RMRULEs to determine which
IRULEs apply to that constituent;but it does not
actuallyapply them until the parser indicatesthat
This buys
all constituents have been found.

example, a noun phrase (NP) can serve


various functions in a clause, including logical
subject (LSUBJ), logical object (LOBJ), surface
subject(SSUBJ),and first NP (FIRSTNP).

*That is, the constituentlabelledLSUBJ must be


interpretable
as a person.

317

efficiency by
rejecting constituent label
assigrments which have no hope of semantic
interpretation,
while deferringthe constructionof
syntactic
an
interpretation until
the
well-formedness
of the entire phrase is verified.
I

of the PHRASE, and may have 0 or more other


SYNTACTIC-CONSTITUENTS
which are Modifiers. The
double arrow or SuperC Cable between PHRASE and
SYNTACTIC-CONSTITUi?~indicates
that every instance
of PHRASE is therebya SYNTACTIC-CONSTITUENT.
The simnlifiedtaxonomv* for our example is
that any CLAUSE
given in Fig: 2-3. This ind*icates
whose Head is the verb "run"

2.2. An exampleof the cascade


simplified example
As
a
of
the
parser-interpreter
interaction,and the use of the
KL-ONE taxonomy of syntactic/semanticshapes in
this interaction,we will briefly describe the
processof parsing the clause "John ran the drill
press." The simplifiedATN grammarwe use for this
exampleis shown in Fig. 2-1.

Figure 2-l: A simplifiedATN


For readers unfamiliarwith KL-ONE, we will
explain three of its major constructsas we note
the informationrepresentedin the simple taxonomy
shown in Fig. 2-2. In E-ONE, Generic Concepts
(ovals in the diagram, boldfaceinhm
represent

Figure 2-3: A simple EL-ONE SyntacticTaxonomy

Figure 2-2: A simple EL-ONE network


description templates, from which individual
descriptionsor IndividualConcepts (shadedovals,
also boldface in text) are formed. In Fig. 2-2,
the
most
general
description
*
SYNTACTIC-CONSTITUENT,
which is specializedby tiz
two descriptions,PHRASE and WORD. All KL-ONE
descriptions are structured objects. The only
structuringdevice of concernhere is the Role. A
Role (drawn as a small square, and underEd
in
text) representsa type of relationshipbetweentwo
objects, such as the relationbetween a whole and
one of its parts. Every Role on a Generic Concept
indicateswhat type of object can fill the Role,
and how many distinct instances of the relation
representedby the Role can occur. The restriction
on fillers of a Role is given by a pointer to a
Generic Concept, and the ntanber of possible
instancesof the Role is shown by a number facet
(indicatedin the form "M < # < NH in thegs
In our diagram we indicatethat every PHRASE has a
WORD associatedwith it which fills the Head Role

318

(independentof tense and person/nunberagreement)


is an example of a RunCLAUSE. There are two
classesof RunCLAUSEsrepresentedin the taxonmy person
(the
LSUBJ
is
a
those
whose
and those whose LSUBJ is a
PersonRunCI.AUSEs),
machine (the MachineRunCLAUSEs). The class of
PersonRunCLAUSEsis again sub-divided, and its
subclassesare RunMachineCLAUSE(in which the LOBJ
must be a machine), RunRaceCLAUSE(in which the
LOBJ is a race), and SimpleRunCLAUSE(which has no
LOBJ).
If we get an active sentence like "John ran
the drill press",the first stage in the parsingis
to PUSH for an NP from the CLAUSE network. For
simplicitywe assume that the result of this is to
parse the noun phrase "John" and produce a pointer
to NPl, an IndividualConceptwhich is an instance
of the Generic pattern PersonNP. This is the
resultof interactionof the parser and interpreter

*To reduce clutter,severalsuperC cables to the


ConceptNP have been left out.

at a lower level of the ATN.

3. IncrementalDescriptionRefinement

Since it is not yet clear what role NPl plays


in the clause (i.e. because the clause may be
active or passive),the parser must hold onto NPl
until it has analyzed the verb. Thus the first
transmissionfrom the parser to the interpreterat
this level is the proposalthat rrrunV(the root of
"ran") is the Head of a CLAUSE. The interpreter
accepts this and returns a pointer to a new
Individual Concept Ckl which it places as an
instanceof RunCLAUSE.

We view the cascadedinteractionof svntactic


semantic
interpretation as
analysis and
implementinga recognitionparadigmwe refer to as
intirementaldescription refinement.
In this
paradigmwe assume we are initiallygiven a domain
of structuredobjects,a space of descriptions,and
rules that determine whic$ descriptionsapply to
each object in the domain. As an example,consider
the domain to be strings of words, the structured
descriptionsto be the parse trees of some grammar,
and say that a parse tree applies to a string of
words if the leaves in the tree correspondto the
sequence of words in the string. In general we
assume each description is structured,not only
describing the object as a whole, but having
ccmponentsthat describe the parts of the object
and their relationshipto the whole as well.

Since the parser has by now determined that


the clause is a simple active clause, it can now
transmitthe proposalthat NPl is the LSUBJ of CLl.
Because NPl is an instance of a PersonNP, the
interpreter can tell that it satisfies the
restrictions on the LSUBJ of one of the
specializationsof RunCLAUSE, and thus it is a
semanticallyplausibleassignment. The interpreter
fills in the LSUBJ Role of CL1 with NPl, and
connectsCL1 to PersonRunCLAUSE,
since that is the
only subConcept of RunCLAUSE which can have a
PersonNPas its LSUBJ.
Finally, the parser PUSHes for an NP,
resulting in a pointer to NP2, an instance of
MachineNP. This is transmittedto the interpreter
Since CL1 is a
as the LOBJ of CLI.
PersonRunCLAUSE,
the taxonomyindicatesthat it can
be either an instance of a RunRaceCLAUSEor a
RunMBCLAUSE,
or a SimpleRunCLAUSE. Since-P2
has been classifieras an instanceof MachineNP,it
is not compatible with being the LOBJ of a
RunRaceCLAUSE(whoseLOBJ must be interpretableas
a race). On the other'handNP2 is compatiblewith
the restrictionon the filler of the LOBJ Role of
RunMachineCLAUSE.
We assume that the taxonomyindicatesall the
acceptablesubcategoriesof PersonRunCLAUSE~Thus
it is only semanticallyplausiblefor NP2 to fill
the LOBJ Role of CL1 if CL1 is an instance of
RunMachineCLAUSE. This being the case, the
interpretercan join CL1 to RunMachineCLAUSEand
fill its LOBJ Role with NP2, creatinga new version
of CL1 whmit
returnsto the parser.
At this point, since there are no more words
in the string, the parser transmits a special
message to the interpreter,indicatingthat there
are no more constituentsto be added to CLI. The
interpreter responds by finding the IRULEs
RunMachineCLAUSE,
from
inherited by
CL1
PersonRunCLAUSE,etc. and using the actions on
those IRULEs to create the interpretationof CLl.
It associates that interpretationwith CL1 and
returns a pointer to CLl, now a fully parsed and
interpretedclause,to the parser.

We consider a situation that correspondsto


left-to-right
parsing. A machine is presentedwith
facts about an object or its parts in some
specifiedorder, such as learning the words in a
string one by one in a left-to-rightorder. As it
learns more propertiesof the object the machine
must determine which descriptionsare compatible
with its currentknowledgeof the propertiesof the
object and its parts.
Incremental description refinement (IDR) is
the processof:
o determining the set of descriptions
ccmpatiblewith an object known to have a
given set of properties,and
o

refining the set of descriptionsas more


propertiesare learned.

More precisely,for every set of propertiesP


= {pl,...,pn)of some object 0 or its parts, there
is an associated set of descriptionsC(P), the
descriptivecover of P. The descriptivecover of P
consists ofthose descriptionswhich might
possibly be applicableto 0, given that 0 has the
that is, the set of
properties PI,-**Ib;
descriptionswhich apply to at least one object
which has all the propertiesin P.
As one learnsmore about some object, the set
of descriptions consistent with that knowledge
shrinks. Hence, the basic step of any IDR process
is to take (1) a set of propertiesP, and (2) its
cover C(P), and (3) some extensionof P into a set
P', and to produce C(P') by removing inapplicable
elements from C(P). The difficultyis that it is
usually impractical, if not impossible, to
represent C(P) extensionally:in many cases C(P)
will be infinite. (For example,until the number
of words in a string is learned, the nLPnberof
parse trees in C(P) remainsinfiniteno matter how
many words in the string are known.) Thus, the

*Actually, the interpreter creates a Generic


subConcept of RunCLAUSE, in order to facilitate
sharingof informationbetweenalternativepaths in
the parse, but we will ignore this detail in the
remainderof the example.

319

*We assume that at least one descriptionapplies


to each object in the domain.

covering set must be represented intensionally,


with the consequence that "removing elements"
becomes an inferenceprocess which determinesthe
intensional representation of C(P1) given the
intensionalrepresentation
of C(P). Note that just
as any element of C(P), representedextensionally,
may be structured, so may the intensional
representation
of C(P) be structuredas well.

it as an IDR process as we have describedabove.


Briefly, we can view the KL-ONE taxonomy of
syntactic/semantic shapes
as
a
set
of
discriminationtrees, each with a root labelledby
some syntactic phrase type. At each level in a
tree the branchesmake discriminations
based on the
properties of some single labelled constituent
(El, such as the LSUBJ of a CLAUSE.

The trick in designing an efficient and


effective IDR process is to choose a synergistic
inferenceprocess/intensional
representationpair.
One example is the use of a discriminationtree.
In such a tree each terminal node representsan
individualdescription,and each non-terminalnode
representsthe set of descriptionscorrespondingto
the terminalsbelow it. Every branch indicatesa
test or discriminationbased on some property (or
properties)of the object to be described. Each
newly learnedpropertyof an object allows the IDR
process to take a single step down the tree, as
long as the properties are learned in an order
ccmpatiblewith the tree's structure. Each step
thus reducesthe set of descriptionssubsumed.

The parser first proposes a phrase type such


as CLAUSE, and the IDR process determines which
tree has a root with that label. That root beccmes
the current active node in the IDR process. All
further refim
isone
within the subtree
dominated by an active node.
As the parser
proposesand transmitsnew LC's to the IDR, the IDR
may respondin one of two ways:

Another IDR process is the operation of a


constraintpropagationsystem. In such a systeman
object is described by a set of nodes, each of
which bears a label chosen from some fixed set.
The nodes are linkedinto a network,and there is a
constraintrelation that specifieswhich pairs of
labels can occur on adjoining (i.e. linked) nodes.
The facts learnedabout an object are either links
betweenpreviouslyknown nodes,or label sets which
specify the possible labels at a single node. A
descriptivecover is simply the cross-productof
some collectionof node label sets. The refinement
operationconsistsof (I) extendingthe analysisto
a new node, (2) removing all incompatiblelabels
from adjacent nodes and (3)
propagating the
effects. Unlike the use of a discriminationnet,
constraint propagation does not require that
informationabout nodes be considered in some -a
priori fixed order.
As mentioned earlier, in the RUS frameworkwe
are attemptingto refine the semantic description
of an utterance in parallel with determiningits
syntacticstructure. The relevant propertiesfor
this IDR process include the descriptions of
variousconstituentsand their functionalrelations
to their matrix (cf. Section 2). Unfortunately,
surfacevariationssuch as passiveforms and dative
movementmake it difficultto assume any particular
order of discovery of properties as the parser
considerswords in a left to right order. However,
the taxonomiclattice of KL-ONE can be used as a
generalizationof a discriminationtree which is
order independent. The actual operation used in
PSI-KLONEinvolvesan extendednotion of constraint
propagationoperating on nodes in the taxonomic
lattice, and thus the resulting system has
interestinganalogiesto both simpler forms of IDR
processes.
The completealgorithmfor the IDR process in
PSI-KLONEis too ccmplex to cover in this paper,
and will be described in more detail in a
forthcomingreport. However, the reader is urged
to returnto the examplein Sec. 2.2 and reconsider

1.

it may reject the LC because it is not


compatiblewith any branch below the
currentlyactive node(s),or

2. it may accept the LC, and replace the


current-58376 node(s) with the (set of)
node(s) which can be reached by branches
whose discriminations
are compatiblewith
the LC.
4. The IDR Process and KnowledgeRepresentation
We-haveidentified
four
critical
characteristics of any general representation
scheme that can support an IDR process in which
descriptions are
structured and
covering
descriptionsare represented intensionally. In
such a scheme it must be possible to efficiently
infer from the representation:
1.

what properties of a structured object


provide sufficient information to
guarantee the
applicability of
a
description to (some portion of) that
object- i.e.,criterialityconditions,

2. what mappings are possible between


classes of relations e.g. how
functional
relationships
between
syntacticconstituentsmap onto semantic
relationships
3.

which pairs of descriptionsare mutually


incompatible- i.e.,cannot both apply to
a single individual

of descriptions
4. which sub-categorizations
are exhaustive- i.e., at least one of
the sub-categories
appliesto anythingto
which the more general description
applies.
Wr analysis of the assumptionsimplicit in
the current implementation
of PSI-KLONEhas led us
to an understandingof the importanceof these four
points in a IDR. By making these four points
explicitin the next implementation
we expect to be
able to deal with a larger class of phenomenathan
the current system handles. In the following
sectionswe illustratethese four points in terms
of the behaviorof the currentversionof PSI-KLONE
and the improvementswe expect to be able to make

320

with more explicit


information

of

types

of

3.

4.1. CriterialityConditions
The point here is an obvious one, but bears
repeating. If a taxonomy is to be used for
recognition,then there must be some way, based on
partialevidence,to get into it at the right place
for the recognition(IDR) process to begin. That
is, for any ultimatelyrecognizablephrase there
must be at least one criteria1 condition,i.e. a
collectionof facts which is sufficientto ensure
the abnlicabilitvof some particulardescris
In the'syntactic"/semantic
taxonomy, the criteria1
conditionis often, for a phrase,the propertiesof
belongingto a particularsyntacticcategory(e.g.,
noun phrase,clause, etc.) and having a particular
lexical item as head. Recallingthe example given
in Section2.2, the evidencethat the input had the
shape of a CLAUSE and had the verb llrunVas its
head constitutedsufficientconditionsto enter the
taxonomy at the node RunCLAUSE - i.e., a
specializationof CLAUSE whose head is filled by
the verb "run". Without the notion of criteria1
properties,we cannot ensure the applicabilityof
any description and therefore have no way of
continuingthe recognitionprocess.

there is an abstractionhierarchy among


slots (the Role hierarchyin KL-ONE),as
well as the more common IS-A hierarchy
among frames (the SUPERC link between
conceptsin KL-ONE),

then the interpreter can make use of this


abstractionhierarchy in answeringquestions from
the parser.
As an example, consider a question from the
parser, loosely translatableas "Can the PP 'on
Sunday' be a PP-modifierof the NP being built
whose head is 'party'?".
Fig. 4-1jo

Figure4-l: A Simple NP Taxonomy


4.2. MappingSyntacticto SemanticRelations
the parser intermittently sends
In RB,
messages to the interpreterasking whether it is
semanticallyplausiblefor a constituentto fill a
The interpreter's
specified functional role.
ability to answer this question ccmes from its
RMRULEs and their organization. This is based on
the assumption that a potential constituentcan
fill some functionalrole in the matrix phrase if
and only if it also fills a semantic role
compatiblewith:
o

that functionalrole

the interpretation
of that constituent

the head of that matrix phrase

the roles filledby the other constituents


of that phrase

We assumethe NP headed by "party"has alreadybeen


classifiedby the interpreteras a CelebrationNP.
As indicated in Fig. 4-l this concept inherits
from EventNP two specializationsof the general
NP
PP-modifier relation applicable to
location-PP-modifier
and time-PP-modifier. Thus
p-modifiers iff it
"on SundaT can be one of%
can be either its location-PFmodifieror its
time-PP-modifier. The next sEtion will discuss
howt&
decisioncan be made. The point here is
that there must be some indication of which
syntactic relations can map onto which semantic
ones, and under what circumstances.An abstraction
hierarchyamong Roles providesone method of doing
so.
4.3.

o other syntactic/semanticproperties of
that phrase and its constituents.
With respectto the first of these points,one
effective way of representing the compatibility
restrictions between syntactic and semantic
relations derives from the fact that each purely
syntacticrelationcan be viewed as an abstraction
of the syntacticpropertiessharedby some class of
semantic relations (i.e., that they have
syntactically
identicalargunents). If
1.

a general frame-likesystem is used to


representthe system'ssyntactic/semantic
knowledge,

2. possiblesyntacticand semanticrelations
are representedtherein as "slots" in a
frame,and

Explicitcompatibility/incompatibility
annotations
As noted above, the semanticinterpretermust
be able to decide if the interpretation
assignedto
the already parsed constituentis compatiblewith
the type restrictions on the argLPnentsof a
semanticrelation. For example,the PP "on Sunday"
can be a PP-modifierof an NP whose Head is "party"
if it isccmpatible with being either a time-PP,
and hence capable of instantiatingthe relation
time-PP-modifier,or a location-PP and hencs
instaziating the relation location-PP-modifier.
ies for formalizing
There are two plausiblestrateg
the somewhatinformalnotion of compatibility:

*In this, as in Fig. 4-1, we assume for


simplicitythat only these two semanticrelations
are consistent with the syntactic relation
-PP-modifierfor an NP whose head is "party".

321

1.

needed to resolve the metonymy, since it would


indicate that "the hamburger" is possibly being
used metonymouslyto refer to some discourseentity
which is both describable by an animate-NP and
associatedwith some (unique)hamburger.

a constituentis judged ccmpatiblewith a


restriction if its syntactic/semantic
shape
(and
hence
interpretation)
guarantees consistency with the type
restrictions,
or

2. it
is
compatible if
its
3-&W
interpretation does
not
guarantee
-inconsistency.
Considerthe problem of rejecting"on Sunday"
as a location-PP-modifier.Conceivablyone could
reject it on thegrounds that "Sunday"doesn'thave
a syntactic/semantic
shape that guaranteesthat it
is a location-NP. This is essentiallythe strategy
followedby the currentversionof PSI-KLONE. More
specifically,the PSI-KLONEsystem searches along
the superC cables of a constituentto find just
those semanticrelationswhich are guaranteedto be
canpatible with the interpretation of the
constituentand matrix.
However, that strategy would have to reject
"birthday present" as being compatible with
apparel-NP (thereby rejecting "Mary wore her
birthdaypresentto New York"),vehicle-NP(thereby
rejecting "Mary drove her birthday present from
Boston to Philadelphia"), animate-NP (thereby
rejecting "Mary fed her birthday present some
Little Friskies"), etc. Thus, we believe that
future systems should incorporate the second
strategy, at least as a fall-back when no
interpretation is found using only the first
strategy. This strategyalso makes it easier for
the system to handle pronouns and other
semantically empty NPs (e.g. "thing" "stuff"
etc.) whose syntactic/semanticshapes'guarantei
almost nothing,but which are compatiblewith many
semanticinterpretations.
The imp1ication here for both language
processingand knowledgereprmesentation
is that:
1.

incompatibility
must be marked explicitly
in the representation,
and

2. the most useful strategyfor determining


compatibilityinvolvesnot being able to
show explicitinccmpatibility.
One caveat and one further observation: this
strategy is not by itself effective in certain
cases of metonymy,which Webster'sdefines as "the
use of the name of one thing for that of another
associatedwith or suggested-byit." For example,
semantics would reject "the hamburger" as the
subjectof,a clause like "the hamburgeris getting
impatient" which might occur in a conversation
between a waiter and a short-ordercook. However,
the taxonomywould be able to provide information

The observation concerns the way in which


semantic interpretationwas done in LUNAR [lOI,
which was to judge semanticcompatibilitysolelyon
the basis of positivesyntactic/semantic
evidence.
A semantic interpretation rule could only be
appliedto a fragmentof a parse tree if the rule's
left-hand side - a syntactic/semantic
template could be matched against the fragment. The only
kinds of semanticconstraintsactuallyused in the
LUNAR templateswere predicateson the head of some
tree constituent-- e.g. that the head of the NP
object of a PP constituentwere of class element,
rock, etc. Given this restriction,LUNAR w-t
be able to handle an utterance like "give me
analysesof alLaninunin NASA's gift to the Royal
Academy",where clearly"gift to the Royal Academy"
is not incompatiblewith rock.
4.4. Explicitmarkingof exhaustive
sub-categorization
in the taxonomy
The algorithm we
have developed for
incrementaldescriptionrefinement requires that
the IDR process be able to distinguishexhaustive
from non-exhaustive sub-categorizationin the
taxonomy
of
syntactic/semantic shapes.
Exhaustiveness
marking plays a role similarto that
played by inclusive or in a logical framework.
That is, it justifies the application of
case-analysis techniques to the problem of
determiningif a proposedconstituentis ccmpatible
with a given syntacticrole. The interpreteris
justified in rejecting a proposed label for a
constituentonly if it has consideredall possible
ways in which it can correspond to a semantic
relation.
Exhaustivenessmarking also make it possible
to infer positive information from negative
informationas was done in the example in section
2.2. There, the interpreterinferred that the
clausewas a RunMachineCLAUSE,
because it was known
to be a PersonRunCLAUSE
and the proposedLOBJ was
incompatiblewith it being a RunRaceCLAUSE. Such
reasoning is justifiedonly if the subcategories
RunMachineCLAUSE,
SimpleRunCLAUSE
and RunRaceCLAUSE
exhaustthe possibilitiesunder PersonRunCLAUSE.
These types of inferencedo not always ccme up
in systemsthat are primarilyused to reason about
inheritedpropertiesand defaults. For example,as
long as one knows that DOG and CAT are both
specializations
of PET, one knows what properties
they inherit from PET. It is irrelevantto an
inheritanceprocess whether there are any other
kinds of PET such as MONKEY, BOA-CONSTRICTORor
TARANTULA.
Many formalisms, including KL-ONE, do not
require the sub-categorization
of a node to be
exhaustive. So there are two optionsvis-a-visthe
way exhaustiveness
can be indicated. A recognition
algorithm can act as if every node were
exhaustivelysub-categorized- a type of Closed
World Assumption L-81- this is essentiallythe way

*If we assume somethinglike "hamburger"being an


instance of definite-food-NP,
which is marked as
incompatiblewith animate-NP,the restrictionon
the subjectof "impatient".

322

operates.
current
system
the
PSI-KLONE
Unfortunately,there are other uses of KL-ONE in
the natural language system in which concepts are
subcategorizedbut it is clear that an exhaustive
If the
subcategorizationhas not been made.
meaning of the links in the representationscheme
is to be well-defined,it must be possible to
non-exhaustive
distinguish exhaustive from
sub-categorization. The implication for both
knowledgerepresentation
and inferenceis that some
clear stand must be taken vis-a-vis the
representation
of exhaustivesub-categorizations.

REFERENCES
[II

PI

L-31

5. Conclusion
The approach we have taken in RUS is midway
betweencompletelydecoupledsyntacticand semantic
processingand the totally merged processingthat
is characteristicof semantic grammars. RUS has
already proven the robustnessof this approach in
several different systems, each using different
knowledge representation techniques for the
semantic ccmponent.
The RUS grammar is a
substantialand general grammar for English,more
extensive than the grammar in the LUNAR system
Althoughthe grammar is representedas an
ClOl.
ATN, we have been able to greatly reduce the
backtrackingthat normallyoccurs in the operation
of an ATN parser, allowing RUS to approach the
performanceof a "deterministic"
parser [21. With
the aid of a "grammarcompiler" [51 this makes it
possibleto achieve parsing times on the order of
.X CPU seconds, on a DEC KLIO, for twenty word
sentences.

[41

[51

Ii61

In this paper we have-focused on the latest


embodimentof the RUS framework in the PSI-KLONE
system -- in particular on the nature of its
cascaded syntactic/semanticinteractionsand the
incremental description refinement process they
support. We believe that what we are learning
about such cascadedstructuresand IDR processesin
building PSI-KLONEis of value for the design of
both natural language systems and knowledge
representation
systems.

Inc., May,
c71

[81

ACKNOWLEDGEMENTS
Cur work on the PSI-KLONEsystem has not been done
in vacua - our colleaguesin this researcheffort
includeEd Barton, Madeleine Bates, Ron Bra&-man,
Phil Cohen, David Israel, Hector Levesque, Candy
Sidnerand Bill Woods. We hope this paper reflects
well on our joint effort.
The authors wish to thank Madeleine Bates,
Danny Bobrow, Ron Bra&man, David Israel, Candy
Sidner, Brian Smith, and Dave Waltz for their
helpfulcommentson earlierversionsof this paper.
Our specialthanks go to Susan Chase,whose gourmet
feasts and general supportmade the preparationof
this paper much more enjoyable than it might
otherwisehave been.

191

Cl01

Cl11

r121

323

Bobrow, R. J.
The RUS System.
-BBN Repom8,
Bolt Beranekand Newnan
Inc,, 1978. .
Bobrow,R. J. & Webber,B. L.
PSI-KLONE- Parsingand Semantic
Interpretation
in the BBN NaturalLanguage
UnderstandingSystem.
In CSCSI/CSEIOAnnual Conference.
7z?xrim,*1980.
Brachnan,R. J.
On the Epistemological
Statusof Semantic
Networks.
In Findler,NicholasV. (editor),Associative
Networks- The Representationand Use of
Knowledgeinomputers, . AcaGi'Gs,
New York,379.
Bract-man,
R. J.
An Introductionto KL-ONE.
In Bra&man, R.J., et al. (editors),Research
in NaturalLanguageUnderstanding,
Annual
Rep31 Aug. 79), pages
--13;116, Bolt Berzek and NekananInc,
Cambridge,MA, 1980.
Burton, R. & Woods, W. A.
A CompilingSystem for AugmentedTransition
Networks.
In COLING76. Sixth InternationalConference
?%7ZiputationalLinguistics,Ottawa,
Canada,June, 1976.
Burton,R.. Seelv Br0wn.J.
SemanticGrammar:A Techniquefor
ConstrwNa?%ral
LanguazInterfaces
to Instructional
Svstems.
BBNTeport 3587, E!ol-?&%%k And Newnan
1977.

Mark, W. S. & Barton,G. E.


The RUSGrammarParsingSystem.
m
3243, Genermrmarch
Laboratories,1980.
Reiter,R.
ClosedWorld Data Bases.
In Gallaire,H. & Minker,J. (editor),Logic
and Data Bases, . PlenumPress, 197T
-Woods, W. A.
TransitionNetworkGrammarsfor Natural
LanguageAnalysis.
CACM 13(10), October, 1970.
Woods, W. A., Kaplan,R. M. & Nash-Webber.B.
The Lunar SciencesNaturalLanguage
'
-InformationSystmnal
Report.
BBN Report 2378, Bolt Beranekandwnan
Inc., June, 1972.
Woods, W. A.
TaxonomicLatticeStructuresfor Situation
Recognition.
In TheoreticalIssues in NaturalLanguage
Processing-rACL amSIGART,
July,
1978.
Woods, W. A.
CascadedATN Grammars.
Amer. J. Computational
Linguistics6(l),
Jan:-Mar.,1980.