Professional Documents
Culture Documents
Ann Copestake
StanfordUniversity,Universityof Cambridgeand Universityof Stuttgart*
aac@csli, stanford, edu
21
reelwastakentohavea basicunderspecified representa-The same polysemypatternsare oftenfoundcross-
tion,whichstatedthatit wasa container artifact with linguistically, but may be accompanied by morpholog-
the purposeof (un)winding, wherethe materialwound icaleffects in onelanguage thatarenotmirrored in the
is leftlargelyunspecified.In a usesuchascotton reel, other.For example, in Englishthe samewordis often
the objectof the (un)winding is specialised. We rep- usedfora plant andforitsfruit (e.g. olive),butinSpan-
resentthisusingPustejovsky’s theoryof lexical seman- ishthereisa gender distinction (e.g.aceituna//aceituno).
tics(e.g.Pustejovsky 1991).Underthisaccount, the Senseextensions oftenseemtofallintofamilies, asin the
representation
of nounsincludes a specification of their examples in (2),wheremeat-grinding andfur/skin grind-
qualiastructure, whichencodes theform,constituency,ingcanbe regarded as conventionalised special casesof
agentiveandtelic(purpose) roles.Thusthetelicrole grinding.
of thebasicsenseofreelwouldbe onlypartially instan-
tiated,butcouldbecomefurther instantiated contextu- (2)a Sandylikesto eatrabbit.
ally.However,physical differences between typesof reel animal -> meat (meat-grinding)
aretreatedas outside thedomain of lexical semantics. b Sandy likes to wear rabbit.
Ouruseof default inheritance allows us to extend this animal -> fur/skin (fur/skin grinding)
styleof representationto examples suchas cloud,which c That stuff on the tarmac looks like rabbit.
is treated
ashaving a basicsensewitha specified formal physical object -> substance (grinding)
role,buta default constitutive role(water vapour).
phrasessuchaa cloudof mosquitos, thedefault is over- However languages differ in what sense extensions are
ridden.Somefurtherexamples of constructional poly- possible: for example, Nunberg and Zaenen (1992) re-
semyare discussed by C+B and Lascarides (thisvol- port that Eskimo has no conventionalised meat-grinding.
ume),includingadjectival interpretation (fastcarver- Finally, some extended senses are established while oth-
sus fasttypist,Pustejovsky and Boguraev, 1993)and ers are not, as illustrated in (3):
logicalmetonymy (Pustejovsky, 1991).Whatall these (3) a That restaurant serves ostrich.
caseshavein commonis thatconventional monotonic animal -> meat
accountscanonlyassigna veryunderspecified represen- b The ham sandwiches has paid his check.
tation,whichis assumed to be instantiated by pragmat- physical object -> associated person
ics.Clearly,without an account of thepragmatic compo-
c [Chester] serves not just country folk, but
nentandtheinterface between it andsyntax/semantics, farming, suburban and city folk too. You’ll
theorieswhichmakethisclaimaredifficult to falsify,
see Armani drifting into the Grosvenor Ho-
and fewformalaccounts havebeenattempted (although
tel’s exclusive (but exquisite) Arkle Restau-
seeHobbset al (e.g.1990)).But forall theseexam-
rant and C+Agiggling out of its streetfront
plea,a pragmatic
explanation runsintodifficulties under
brasserie next door. (Guardian Weekly)
standardassumptions aboutthelimitations on the type
manufacturer -> product +
of informationthe pragmatic component has accessto.
clothes -> wearer
Thisis discussed
in thereferences above. Thealternative
approachis to assumea relatively
richlexicalsemantic The essential points of our analysis are that sense ex-
representation,suchasqualiastructure,butthishasto tensions are treated by lexical rules which are similar in
be donein a representational frameworkwhichallows all respects to rules which encode derivational morphol-
defaults, in orderto allowpragmaticsto override
the 1ogy except for the change in orthography/phonology.
lexicallyinstantiateddefaults.
Lascarides
(thisvolume) Sense extension rules can inherit from one another. For
explains howthe lexical/ pragmatic interfacemaybe example, the meat-grinding and fur/skin grinding rules
formaiised--sinceshealsogivesthemoresalientdetails illustrated above can be regarded as subtypes of the more
of the C+Baccounts of logical
metonymy andadjectival general grinding rule. Because we use a fine-grained lex-
interpretation,I willnotrepeatthemhere. ical semantic representation, we can be quite specific
about the class of lexical items to whicha rule is intended
3 Sense extension to apply. Wesketch the grinding and meat-grinding rules
in Figure 1 (see C+Bfor a full description). Note that
In previouspaperswe havearguedfor an approachto the semantic effect of grinding is captured by an operator
examplessuchas thosein (I)whichtreatsthemas ex- which applies to the predicate, while the specialisation of
tensions
froma basesense.
the meaning in meat-grinding is indicated in the qualia
(i) a He boughttwo beers. structure.
comestible substance-> conventional
por- The motivation for making the distinction between
tion(portioning) constructional polysemy and sense extension is discussed
b He dranka bottleof whisky.
container-> contents t The term lexical rule has come to seem something of
a misnomer,but this is rather of peripheral interest here. I
Senseextensions may havesyntactic/grammatical ef- will also ignore the issues whicharise whenformalisinglexical
fects:e.g.portioning(la)convertsmassto count.Some rules in a constraint based framework:for current purposes
senseextensionsarecomparable in semantic effectto wecantreatlexical rulesasbeingsimilar to non-branching
morphologicalprocesses.The ’container-> contents’ phrasestructure rulesor schemata (cfRiehemann’s (1993)
examplein (ib)is paralleled
by suffixationwith-ful. treatment of derivational morphology within HPSG).
22
grinding < lexical-rule sume that speakers choose well-attested high-frequency
forms to realise particular senses and listeners choose
ORTH = [] oITrH = []
SYN ffi noun-cat -+ SYN = noun-cat well-attested high-frequency senses when faced with am-
SEM PRED = [~] SEMPREDffi grinding’(0J]) biguity, then muchof the ’semi-productivity’ of lexical
QUALIA = phylical QUALIA = physical
rules is predicted. For instance, we wouldpredict that in
meat-grinding < grinding the ’null’ or a neutral context 2 (4a) will be interpreted
[ QUALIA. animal ]--~[ QUALIAffi edible.~ubstance ] as rabbit meat, and (4b) will be interpreted as animals.
Less frequent but attested senses could be chosen when
Figure 1: Grinding and meat-grinding dictated by the context, as in (4c).
(4) a John prefers rabbit.
b John wants three rabbits.
extensively in C+Band we will not repeat it here, but
there are several reasons why simple inheritance is not c The diners ordered three rabbits.
adequate for lexical rule representation. Rules can feed Probability also plays a role in the application of lex-
each other: (3c) appears to be a case of this, and ical rules in novel usage. Under the current proposal,
a context such as (4c), below three rabbits could mean lexical rules are somewhatlike ’redundancy’ rules in that
three portions of rabbit meat. Furthermore in order to they can be used to construct appropriate signs for in-
describe the semantic effect of lexical rules it seems nec- stitutionalised senses of a word form, which will have a
essary to postulate operators such as grinding’, which non-zero probability in the lexeme entry. However, in
cannot be satisfactorily represented using simple inher- the situation where an interpretation for a novel usage
itance, because there is no (substantive) commoncore is called for, an assessment of the relative probability of
meaning to the senses which can be specialised. Simple possible lexical rules would provide a means for adopt-
inheritance is knownto be inadequate for derivational ing the most likely ’analogous’ interpretation. For in-
morphology, for similar reasons, and since we would like stance, interpreting an examplessuch as (5), the listener
to use the same mechanismfor both derivation and sense who had not experienced examples of any variant of the
extension this further argues for the use of lexical rules. grinding rule with guinea pig would choose the rule with
the highest probability given the semantic type of the
3.1 Seml-productivity noun.
Lexical rules must be able to generate nonce senses to (5) John prefers guinea pig
deal with examples such as those in (3). But, at least
from a computational perspective, this leads to a con- The probability of a lexical rule can be derived by com-
siderable problem since lexical polysemy then becomes paring the number of lexemes to which the rule could
intractable. It is not even clear whether the lexicon can apply where that sense is unattested, to those for which
be regarded as finite, since sense extension rules can it is attested. The following formulation is due to Ted
feed one another (again this is also an issue in deriva- Briscoe:
tional morphology). We described an approach to re-
Prob(lexical-rule) EN Prob(leo [wf0)
stricting productivity in C+Bwhich I want to elaborate N
on here. Weargued that in the absence of other factors,
language users utilise frequency information to resolve (where N = number of lexical entries which match the
indeterminacies in both generation and interpretation. lexical rule input and leo and wfo denote the output
Such a strategy is compatible with and may underlie the lexical-entry and output word-form, respectively).
Gricean Maximof Manner, in that language will be more Weregard our use of probabilities as being consistent
easily interpretable if there is a tacit agreement not to with Bauer’s claim that accounting for semi-productivity
utilise abnormal or rare means of conveying particular is an issue of performance, not competence (Bauer
messages. Wecan model this aspect of language use as 1983:71f). The effect of the use of probabilities is that
a conditional probability that a word form will be used unattested senses will never be assumed unless no at-
in a specific sense: Prob(lexical-entry I word-form). tested sense is possible, thus allowing interpretation of
assume that probabilities are associated with all estab- examples such as those in (3), while avoiding overgener-
lished forms, regardless of whether they can be treated ation. This improves on the control principle suggested
as derived from other senses, and we make no claim that in Copestake (1992), that lexical rules should only be ap-
a derived sense will necessarily be less frequent than a plied if no interpretation was applicable which did not
basic one; for examplefor turkey our intuition is that the involve a lexical rule, since it allows for cases such as
meat sense is more frequent than the basic animal sense. turkey without further stipulation. The two other con-
For example, a word form such as rabbit can be as- trol effects suggested in Copestake (1992) are both also
sociated with an entry like that illustrated in Figure 2, superseded by the current proposal. One of these was to
in which meat grinding is shown to be (hypothetically) allow for blocking, which is discussed below. The other
more probable than grinding, meat grinding followed by 2FollowingLascarides (this volume) we can give a formal
portioning, or fur/skin grinding. The attribute LRSas- definition of what we meanby neutral context -- it corre-
sociated with the lexeme for rabbit records which combi- sponds to a situation wherethere are no applicable pragmatic
nations of lexical rules have been attested with what fre- rules whichaffect the interpretation of the lexical item under
quency in the experience of the language user. If we as- consideration other than those licensed by the sentence itself.
23
"I4Jxeme
I
F lex-count-noun
[ ORTH = rabbit
SIGN =. ]CAT = noun-cat
BASE a
L SEM ~ obj-noun-formula
PROB = 0.4
LI~S I
([ PIJ’3LE
PROB
"r~i~’
= 0.0S J’
l
[ -
RULE
PROB = -.--
t-Krindi-,K
]r -
’[
RULE "est’KrdK’~’P°rti°nln’
PROB = 0.15 ’
[ PROB =
24
examples. Thus there are some linguistic constraints on tablished, but nevertheless fall into somegeneral classes.
compoundinterpretation, which have to be represented. Someof the established examples are lexicalised, that is
Even with a pragmatic account, some compounds, they would have lexical entries which would augment or
such as garbage man, must be explicitly listed as estab- override the productive interpretation: deckchair, for ex-
lished, because in the ’null-context’ the established in- ample, is no longer particularly associated with a ship’s
terpretation is the only possible one. So either the prag- deck.
matic component has to contain lexical information 3 or The structure below shows the result of instantiating
the lexicon has to contain some compounds with their the schema in Figure 4 with wickerwork Chair (ignoring
established meaning. Sentences containing such com- the substructure in wickerwork).
pounds would be ambiguous because the corresponding
productively generated underspecified compoundwould SEM= wioAcerwork(l~]) ^ chair[’6q ^ /mad~-or-*ub.tanc*( [’6"I, [~])
QUALIA ffi artefact
still have to be available. Andif established compounds
are listed in the lexicon, then any generalisation about In normal contexts, this interpretation will stand. How-
the behaviour of classes of compounds should also be ever, since it is defeasible, it can be pragmatically over-
accessible to the lexicon, since many established com- ridden along the same lines as the examples discussed
pounds have an interpretation that belongs to one of the by Lascarides. Thus in a context such as (9), an alter-
standard patterns. native interpretation would be found, since the default
These observations make noun-noun compounds a interpretation is contradicted by the context.
good candidate for the use of defaults which persist be- (9) At school, everyone worked on crafts in groups
yond the lexicon, where a lexically instantiated default round a big table, sitting on brightly coloured
interpretation can be overridden by subsequent prag- chairs. To make sure everyone could reach the
matic processing, along broadly similar lines to the treat- materials, the groups used particular chairs:
ment of logical metonymyin C+Band Lascarides (this the wickerworkchairs were red plastic, for ex-
volume). A general schema for endocentric compound ample.
interpretation is shownin Figure 3, with an underspeci-
fled predicate, R, relating the indices of the constituents. There are many elaborations that could be made to
Most compoundswill instantiate one or more of the sub- this sketch. As it stands apple juice chair wouldbe given
schemata which inherit from this schema with the pred- the default interpretation
icate relating the parts of the compoundmarked as per-
sistently default. An example of a more specific schema chair(y) A apple-juice(z) A made-of-substance(y,
is shown in Figure 4. The slash notation indicates a de- This would be overridden by pragmatics, but it could
fault value, i.e. the schemadefeasibly specifies that the be lexically excluded since physical entities are subcate-
compounding predicate is made-of-substance. gorised according to physical state, so the rule could be
It should be apparent that under this account it does set up to exclude solid entities being madeof liquids, for
not make sense to claim that compoundshave some fixed example. But I make no claim for the adequacy of this
numberof possible interpretations, since somepredicates particular classification of compoundssince myaim here
are supertypes of more specific predicates, and there is is just to illustrate that the use of the persistent default
not necessarily any limit on specialisation of schemata. mechanismmakes a classification possible in principle.
Multiple schemata may apply to a particular instance of
a compound.The account of semi-productivity of lexical
rules described in the previous section, can be extended 5 Conclusion
to compoundinterpretation, however, so massive ambi- In this paper, I have described an approach to repre-
guity is avoided. senting polysemy, which appears to have considerable
In Figure 5, I tabulate some examples of com- promise for building LKBs, since by using these tech-
pounds taken from a newspaper corpus to illustrate this niques we can get away from the ’dictionary’ notion of
approach.4The sample was just collected for illustrative a fixed number of discrete word senses. Our use of de-
purposes and is much too small (about 100 tokens) fault inheritance and lexical rules gives the lexicon some
draw any firm conclusions, but it is worth noticing that generative capacity with which to interpret novel usages.
the only example which I found that should probably The approach to defaults described here has been imple-
be given a completely underspecified interpretation was mented in the ACQUILEXLKB system. From a com-
Mastermind chair. 5 Most of the other examples are es- putational perspective, the main difficulty in applying
25
compound-noun < binary-rule
lex-noun
Ol~rH = IT], ff]
SYN = noun-cat
SEM= [I]^~^a([]],[]])
QUALIA = [] nomqualia
--~
ORTH = []
SYN = noun-cat
SEM= [] P([~)
QUALIA ffi nomqualia
,
I
ORTH = []]
SYN = noun-cat
the approach to semi-productivity is collecting the prob- Hobbs, J.R., M. Stickel, D. Appelt and P. Martin
abilities, since this requires disambiguation of usages in (1990) ’Interpretation as Abduction’, Technical Note
corpora. This might be feasible, despite the rather dis- No. 499, Artificial Intelligence Centre, SRI Interna-
appointing results of most reported work on sense dis- tional, Menlo Park, CA.
ambiguation since probabilities do not need to be ex- Lascarides, A., E.J. Briscoe, N. Asher and A. Copestake
act, and the sense distinctions we are interested in are (in press) ’Order Independent and Persistent Typed
quite coarse grained compared to those made in most Default Unification’, Linguistics and Philosophy,
learners’ dictionaries. Furthermore Harley (1994) sug- Lascarides, A. and Copestake, A. (1995) ’A Logic
gests that better sense disambiguation is possible, given for Order Independent Typed Default Unification’,
a good lexical database. However,this remains as future ACQUILEX-IIworking paper.
work. Further work would also be required to provide a Levi, J. (1978) The syntax and semantics of complex
computational account of the lexical / pragmatic inter- nominals, Academic Press, NewYork.
face, and a tractable implementation of the pragmatic Nunberg, G.D. and A. Zaenen (1992) ’Systematic poly-
component. semy in lexicology and lexicography’, Proceedings of
Euralex92, Tampere, Finland.
References Pustejovsky, J. (1991) ’The generative lexicon’, Compu-
tational Linguistics, 17(4), 409-441.
Bauer, L. (1983) English word-formation, Cambridge Pustejovsky, J. and B. Boguraev (1993) ’Lexical knowl-
University Press, Cambridge, England. edge representation and natural language process-
Briscoe, E.J., A. Copestake and A. Lascarides (in press) ing’, Artificial Intelligence, 63, 193-223.
’Blocking’ in P. St. Dizier and E. Viegas (ed.), Com- Riehemann, S. (1993) ’Word formation in lex.ical type
putational Lexical Semantics, Cambridge University hierarchies’, M.Phil thesis, University of Tiibingen,
Pr~. Germany.
Copestake, A. (1992) ’The representation of lexical se-
mantic information’, Doctoral dissertation, Univer-
sity of Sussex, Cognitive Science Research Paper
CSRP 280.
Copestake,A. and E.J. Briscoe(in press)’Semi-
Productive Polysemyand SenseExtension’, Journal
of Semantics,
Harley, A. (1994) ’Cambridge Language Survey: Seman-
tic Tagger’, ACQUILEX-IIworking paper No 39:
available from ¢±de@cup.cam. a¢. uk.
26