1

Quantifying aspects ofantonym canonicity in English

and Swedish: textual andexperimental

Carita Paradis, Växjö UniversityCaroline Willners, Lund UniversitySimone Löhndorf, Lund UniversityLynne Murphy, University of Sussex

The Comparative Lexical RelationsGroup

Ongoing project on English, Swedish andJapanese antonyms along two lines ofinvestigation

Antonym canonicityDiscourse functions of antonyms

The Comparative Lexical Relations Group

Aim

Propose a corpus-based method as apossible source for cross-linguisticinvestigations in general andinvestigations of canonicity in particularReport on corpus results and elicitationexperiments

Issues

Are there two distinct types of antonyms oris there a cline from strongly canonicalpairings to weak pairings?Why do most people consider pairs suchas good-bad and long-short as betterantonyms than cold-scorching anddisturbed-fine?

dry wet

parched

wetdry

soggy

anhydrous

sere

arid

watery

damp

moist

humiddried-up

Direct and indirect antonyms as in WordNet

We question this view of a strict dichotomyof direct and indirect antonyms andpropose a continuum from perfect throughgood and less good antonyms to hardlyantonymic at all.

2

Working definition

Canonical antonyms are pairs of words inbinary semantic opposition associated byconvention as well as by semanticrelatedness (e.g. wide/narrow). The notionof canonical antonymy is different fromsemantic opposition in which themeanings are incompatible, but the wordsare not necessarily conventionally paired(e.g. cold/scorching, calm/nervous).

Psycholinguistic investigationsantonyms, canonical as well as non-canonical, tend toelicit each other in psychological tests such as free wordassociation (Deese 1965, Charles and Miller 1989)people are faster at recognizing canonical opposites asantonyms than non-canonical opposites (Herrmann et al.1979, Charles et al. 1994)Charles, Reed & Derryberry (1994) found that canonicalantonym recognition was not affected by the distancebetween members of the pair, while distance in non-canonical antonyms delayed reaction timescanonical antonyms prime each other more strongly thannon-canonical opposites (Becker 1980)

Textual investigationsmembers of canonical pairs co-occur withinsentences at higher than expected rates(Justeson and Katz 1992)they co-occur in sentences significantly moreoften than other potentially antonymous wordpairs (Willners 2001)knowing antonym pairs is not just a matter ofknowing set phrases in which they occur, like thelong and the short of it or neither here nor there.Instead, the same pairs occur in a range ofdifferent contexts and functions (Muehleisen1997, Jones 2002, Jones et al 2005)

We use both experimental and textualmethods to gain insights into the nature ofantonymy as an organizing lexico-semantic principle

Design of canonicity study

Selection of test items from sententiallyco-occurring antonyms in textElicitation experimentsJudgement experimentsAntonym co-occurrence patterns in fixedconstructions using web-as-corpus

Selection of test items

smal-tjockthin-thickTHICKNESS

dålig-brabad-goodMERIT

smal-brednarrow-wideWIDTH

långsam-snabbslow-fastSPEED

liten-storsmall-largeSIZE

svag-starkweak-strongSTRENGTH

ljus-mörklight-darkLUMINOSITY

3

Princeton WordNet

We collected all the synonyms of the 14adjective antonyms from WordNet andthereby got a set of synonyms for each ofthe seven pairs of antonyms. We ran theseven sets of words through the BNC in allpossible constellations in search forsentential co-occurrence (same methodfor Swedish data)

BNC searches

0.00005.9139293596707rapidfast

0.00001.04502239201066suddengradual

0.00000.782195435760tediousslow

0.00009.660916357606707slowfast

P-value

ExpectCo

CoN2N1Word 2Word 1

Test items from the BNC

7 pairs of canonical antonyms at p<10-4

(light-dark, weak-strong etc.)14 pairs of antonyms at p<10-4 (two pairsper dimension)14 pairs of synonyms at p<10-4 (two pairsper dimension)7 pairs of unrelated pairs at p<10-4 (onepair per dimension)

SPEED dimension items

unrelatedhot – smooth

synonymsfast – rapid

synonymsslow – dull

antonymsgradual – immediate

antonymsslow – sudden

(canonical) antonymslow – fast

Test items from Hermann et al.

Hermann et al. had 100 hundred testpairs. We used 11 of their pairs (everysixth gradable adjective pair in their listfrom more canonical through less stronglyantonymical to not related)

Hermann et al’s

glad – irritated

daring - sickhard – yielding

bold – civildisturbed – calm

delightful – confusedtired – alert

nervous – idleimmaculate – filthy

sober – excitedbeautiful -ugly

4

Test items

The English test set consists of 85randomized unique words

The Swedish test set consists of 77randomized unique words

Participants

50 native English participants, 36 womenand 14 men between 19 and 88 years ofage50 native Swedish participants, 25 womenand 25 men between 20 and 70 years ofage

InstructionsYou are going to be given a list with 85 English words. For eachword write down the word that you think is the best opposite for it inthe blank line next to it.Don’t think too hard about it -write the first opposite that you think of.There are no ‘wrong’ answers.Give only one answer for each word.Give opposites for all the words, even when the word doesn’t seemto have an obvious opposite.Don’t use the word not in order to create an opposite phrase. Youranswer should be one word.

Example: The opposite of MASCULINE is _______________ You might answer feminine.

Our predictions are that

there is total agreement across speakersof a language on the pairings of the sevenpairs representing the dimensions, and theother pairs will be less than totalagreement in a gradient, sloping fashion the weaker the degree of canonicity themore responses the test items will yield

To the results… Total agreement

Given X all participants said Y

weak > strong (50)bad > good (50)beautiful > ugly (50)

5

Total agreementGiven X all participants said Y

bra <-> dålig (50)liten <-> stor (50)ljus <-> mörk (50)svag <-> stark (50)

Directionality

Are pairs of antonyms symmetrical ineliciting one another?

If they do,

it could be a sign of canonicity which mayhave several converging reasonsfrequency, monosemy, symmetry, binaryintrinsicness or contextual generality.

If they don’t,

it could be a sign of lower degree ofcanonicity in one direction or in bothdirections due to infrequency, polysemy,asymmetry, non-intrinsic binarity orcontextual specificity.

6

Directionality

Black > white, colour

White > black, dark

Monosemy

Ugly > beautiful, pretty, attractive

Beautiful > ugly

Monosemy

slow - rapid

slow - fast

Polysemy

Good > bad, evil

Evil > good, kind, angelic,pure

Mediocre > outstanding,excellent, exceptional,brilliant, amazing,good…(19)

Polysemy

Light – heavy

Light – dark

Non-intrinsicnessAbundant > rare, scarce, sparse,little, lacking, disciplined, few, limited, needed, none, meagreplentiful, sparing, threadbare

Rare > common, commonplace,ubiquitous, frequent, plentiful,well-known

7

fat > lean, slim, thin,skinny, thick, wrong

lean > fat, fatty, large,plump, support, stocky,wide

slim > fat, broad, big,chubby, wide, large,obese, plump, roundthin > fat, thick,overweight, wide

thick > thin, clever, fine

fine > thick, coarse, badbold, dull, wide, blunt,clumsy, cloudy, mad,ok,wet, unwell

Cluster analysis: English data

16 antonym pairs: strong bidrectionality

13 antonym pairs: weaker or skewedbidirectionality

120 antonym pairs: more disagreementacross participants or strongly skewedresponses

Cluster 1: strong bidirectionality

rapid – slowfast – slow

narrow – wideenormous – tiny

weak – stronglight – darkblack – white

thin - fatlarge – smallbig – small

thick - thinheavy – lightbeautiful – ugly

soft – hardfilthy – cleanbad – good

Cluster 2: weaker or skewedbidirectionalilty

sick - healthygood -evil

narrow - broadfat - slim

tough - tenderhuge - tinydull - bright

thick - finehard - easydark - pale

strong - feeblegradual - suddenbright - gloomy

Cluster 3: examples of weakrelations

bold – weakconfused – finedelicate – robustenormous – slightdark – paleevil – pure

8

Cluster analysis: Swedish data

8 antonym pairs: strong bidirectionality

15 antonym pairs: weaker or skewedbidirectionality

118 antonym pairs more disagreementacross participants or strongly skewedresponses

Dimensions

All the lexical items used for the searchesare strongly bidirectional pairs. That’s notthe case for Swedish where WIDTH is inCluster 2 (instead tired/alert are in Cluster1).

Summary

Proposal of principled corpus-basedmethod for the study of antonyms that canbe used for cross-linguistic investigationsusing experiments or corpora.Some (inconclusive) results of antonymelicitation in English and Swedish to bematched with judgment experiments andantonym constructions in text.

Conclusions

Are there two types of antonyms or isthere a canonicity scale?Textual evidence for the seven dimensionsas a distinct type.Psycholinguistic evidence through clusterspointing towards a cline.

Conclusions

Why are there differences?The stronger pairs are frequent,symmetrical, intrinsically binary (salientdimensions), contextually general andelicit one antoher bidirectionallyBUT, quite a few of the words in Cluster 1are in fact polysemous: good, heavy, light,hard, thick

Thanks for your attention