Investigating adjective denotation and collocation Ann Copestake Computer

Investigating adjective denotation and collocation Ann Copestake Computer Laboratory, University of Cambridge

Outline introduction: compositional semantics, GL and semantic space models. denotation and collocation n distribution of `magnitude’ adjectives n hypotheses about adjective denotation and collocation n semi-productivity n

Themes semi-productivity: extending paper in GL 2001 to phrases n statistical and symbolic models interacting n generation as well as analysis n computational account n

Different branches of computational semantics n compositional semantics: capture syntax, (some) close-class words and (some) morphology n n n lexical semantics, e. g. , n n every x [ dog’(x) -> bark’(x)] large coverage grammars as testbed for GL (constructions, composition, underspecification) GL (interacts with compositional semantics) Word. Net meaning postulates etc semantic space models, e. g. , n n n LSA Schütze (1995) Lin (multiple papers), Pado and Lapata (2003)

semantic spaces n n n acquired from corpora generally, collect vectors of words which co-occur with the target more sophisticated models incorporate syntactic relationships dog bark house cat dog - 1 0 0 bark 1 - 0 0

Semantic space models and compositional semantics? n n do spaces correspond to predicates in compositional semantics? e. g. , bark’ attractions n n n problems n n automatic acquisition similarity metrics, priming fuzziness, meaning variation, sense clustering statistical approximation to real world knowledge? (but fallacy with parse selection techniques) classical lexical semantic relations (hyponymy etc) aren’t captured well can’t do inference sensitivity to domain/corpus role of collocation?

Denotation: assumptions n Truth-conditional, logically formalisable (in principle), refers to `real world’ (extension) n n Not necessarily decomposable: natural kinds (dog’ – canis familiaris), natural predicates Naive physics, biology, etc Computationally: specification of meaning that interfaces with non-linguistic components Selectional restrictions? n bark’(x) -> dog’(x) or seal’(x) or. . .

Collocation: assumptions n Significant co-occurrences of words in syntactically interesting relationships n n `syntactically interesting’: for this talk, attributive adjectives and the nouns they immediately precede `significant’: statistically significant (but on what assumptions about baseline? ) Compositional, no idiosyncratic syntax etc (as opposed to multiword expression) About language rather than the real world

Collocation versus denotation n Whether an unusually frequent word pair is a collocation or not depends on assumptions about denotation: fix denotation to investigate collocation Empirically: investigations using Word. Net synsets (Pearce, 2001) Anti-collocation: words that might be expected to go together and tend not to n n e. g. , ? flawless behaviour (Cruse, 1986): big rain (unless explained by denotation) e. g. , buy house is predictable on basis of denotation, shake fist is not

Collocation and denotation investigations n n can this notion of collocation be made precise, empirically testable? assumptions about denotation determine whether something is a collocation semantic space models will include collocational effects initial, very preliminary, investigations with magnitude adjectives n n attributive adjectives: can get corpus data without parsing only one argument to consider

Distribution of `magnitude’ adjectives: summary n n n some very frequent adjectives have magnituderelated meanings (e. g. , heavy, high, big, large) basic meaning with simple concrete entities extended meaning with abstract nouns, non-concrete physical entities (high taxation, heavy rain) n n extended uses more common than basic not all magnitude adjectives – e. g. tall nouns tend to occur with a limited subset of these extended adjectives some apparent semantic groupings of nouns which go with particular adjectives, but not easily specified

Some adjective-noun frequencies in the BNC number proportion quality problem part winds rain large 1790 404 0 10 533 0 0 high 92 501 799 0 3 90 0 big 11 1 0 79 79 3 1 heavy 0 0 1 2 198

Grammaticality judgments number proportion quality problem large * high heavy ? * big ? ? * * part ? winds rain * * *

More examples impor tance success majority number proport ion quality role problem part winds support rain great 310 360 382 172 9 11 3 44 71 0 22 0 large 1 1 112 1790 404 0 13 10 533 0 1 0 high 8 0 0 92 501 799 1 0 3 90 2 0 major 62 60 0 0 7 0 272 356 408 1 8 0 big 0 40 5 11 1 0 3 79 79 3 1 1 strong 0 0 2 0 0 1 8 0 3 132 147 0 heavy 0 0 1 2 4 198

Judgments impor tance success majority number proport ion quality role problem part great large ? high ? * major * ? ? ? * ? winds ? strong ? ? * * * heavy ? * * rain ? * * * ? big support ? ? * * * ?

Distribution n n Investigated the distribution of heavy, high, big, large, strong, great, major with the most common co -occurring nouns in the BNC Nouns tend to occur with up to three of these adjectives with high frequency and low or zero frequency with the rest My intuitive grammaticality judgments correlate but allow for some unseen combinations and disallow a few observed but very infrequent ones big, major and great are grammatical with many nouns (but not frequent with most), strong and heavy are ungrammatical with most nouns, high and large intermediate

heavy: groupings? magnitude: dew, rainstorm, downpour, rainfall, snowfall, snow, shower: frost, spindrift: clouds, mist, fog: flow, flooding, bleeding, period, traffic: demands, reliance, workload, responsibility, emphasis, dependence: irony, sarcasm, criticism: infestation, soiling: loss, price, cost, expenditure, taxation, fine, penalty, damages, investment: punishment, sentence: fire, bombardment, casualties, defeat, fighting: burden, load, weight, pressure: crop: advertising: use, drinking: magnitude of verb: drinker, smoker: magnitude related? odour, perfume, scent, smell, whiff: lunch: sea, surf, swell:

high: groupings? magnitude: esteem, status, regard, reputation, standing, calibre, value, priority; grade, quality, level; proportion, degree, incidence, frequency, number, prevalence, percentage; volume, speed, voltage, pressure, concentration, density, performance, temperature, energy, resolution, dose, wind; risk, cost, price, rate, inflation, taxation, mortality, turnover, wage, income, productivity, unemployment, demand magnitude of verb: earner

heavy and high 50 nouns in BNC with the extended magnitude use of heavy with frequency 10 or more n 160 such nouns with high n Only 9 such nouns with both adjectives: price, pressure, investment, demand, rainfall, costs, concentration, taxation n

Basic adjective denotation with simple concrete objects: high’(x) => zdim(x) > norm(zdim, type(x), c) heavy’(x) => wt(x) > norm(wt, type(x), c) where zdim is distance on vertical, wt is weight (measure functions, MF) norm(MF, class, context) is some standard for MF for class in context (high’ also requires selectional restriction – not animate)

Metaphor n Different metaphors for different nouns (cf. , Lakoff et al) n n n Empirical account of distribution? n n n `high’ nouns measured with an upright scale: e. g. , temperature: temperature is rising `heavy’ nouns metaphorically like burden: e. g. , workload: her workload is weighing on her predictability of noun classes? high volume? high and heavy taxation adjective denotation for inference etc? via literal denotation? Discussed again at end of talk

Possible empirical accounts of distribution Difference in denotation between `extended’ uses of adjectives Grammaticized selectional restrictions/preferences Lexical selection 1. 2. 3. • stipulate Magn function with nouns (Meaning. Text Theory) Semi-productivity / collocation 4. • plus semantic back-off

Computational semantics perspective Require workable account of denotation: not too difficult to acquire, not over-specific n Require account of distribution for generation n Robustness and completeness n Can’t assume pragmatics / real world knowledge does the difficult bits! n

Denotation account of distribution Denotation of adjective simply prevents it being possible with the noun. n heavy and high have different denotations heavy’(x) => MF(x) > norm(MF, type(x), c) & precipitation(x) or cost(x) or flow(x) or consumption(x). . . n (where rain(x) -> precipitation(x) and so on) n But: messy disjunction or multiple senses, open-ended, unlikely to be tractable. n n n e. g. , heavy shower only for rain sense, not bathroom sense Not falsifiable, but no motivation other than distribution. Dictionary definitions can be seen as doing this (informally), but none account for observed distribution.

Selectional restrictions and distribution n n Assume the adjectives have the same denotation Distribution via features in the lexicon n n n e. g. , literal high selects for [ANIMATE false ] approach used in the Lin. GO ERG for in/on in temporal expressions grammaticized, so doesn’t need to be determined by denotation (though assume consistency) can utilise qualia structure Problem: can’t find a reasonable set of cross-cutting features! Stipulative approach possible, but unattractive.

Lexical selection MTT approach n noun specifies its Magn adjective n n n in Mel’čuk and Polguère (1987), Magn is a function, but could modify to make it a set, or vary meanings stipulative: if we’re going to do this, why not use a corpus directly?

Collocational account of distribution n n all the adjectives share a denotation corresponding to magnitude (more details later), distribution differences due to collocation, soft rather than hard constraints linguistically: n n n adjective-noun combination is semi-productive denotation and syntax allow heavy esteem etc, but speakers are sensitive to frequencies, prefer more frequent phrases with same meaning cf morphology and sense extension: Briscoe and Copestake (1999) blocking (but weaker than with morphology) anti-collocations as reflection of semi-productivity

Collocational account of distribution n computationally, fits with some current practice: filter adjective-noun realisations according to n-grams (statistical generation – e. g. , Langkilde and Knight) n use of co-occurrences in WSD n n back-off techniques

Collocational vs denotational differences high heavy Denotation difference low Collocation difference

Back-off and analogy n n back-off: decision for infrequent noun with no corpus evidence for specific magnitude adjective based on productivity of adjective: number of nouns it occurs with n n default to big back-off also sensitive to word clusters n n n e. g. , heavy spindrift because spindrift is semantically similar to snow semantic space models: i. e. , group according to distribution with other words hence, adjective has some correlation with semantics of the noun

Metaphor again extended metaphor idea is consistent with idea that clusters for backoff are based on semantic space n words cluster according to how they cooccur n n n e. g. , high words cluster with rise words? but this doesn’t require that we interpret high literally and then coerce

More details: denotation of extended adjective uses n mass: e. g. , rain, and some plural e. g. , casualties n n n cf much, many inherent measure: e. g. , grade, percentage, fine other: e. g. , rainstorm, defeat, bombardment n n attribute in qualia has Magn – heavy rainstorm equivalent to storm with heavy rain also heavy drinker etc

More details n n n Different uses cross-cut adjective distinction and domain categories Want to have single extended sense and some form of co-composition Further complications: nouns with temporal duration n heavy rain – not the same as persistent rain heavy fighting but heavy drinking how much of this do we have to encode specifically?

Connotation n heavy often has negative connotations heavy fine but not ? heavy reward etc n heavy taxation versus high taxation n n consistent with the semantic cluster / extended metaphor idea

Necessary experiments n n n None of this is tested yet! Specify denotation, check for accuracy Implement semi-productivity model with back -off Determine predictability of adjective based on noun alone Extension to other adjectives? Magnitude adjectives may be more lexical than others.

Conclusions Testing collocational account of distribution requires fixing denotation n Magnitude adjectives: assume same denotation n more complex denotations would need different experiments Semi-productivity at the phrasal level n Back-off account is crucial

Some final comments denotation, selectional restriction, collocation: choice between mechanisms? n ngrams for language models for speech recognition n variants of semantic space models that are less sensitive to collocation effects? n n can we `remove’ collocation?