This corpus study addresses the question of the nature and the structure of antonymy and synonymy in language use, following automatic methods to identify their behavioral patterns in texts. We examine the conceptual closeness/distance of synonyms and antonyms through the lens of their DOMAIN instantiations.
Using data from Wikipedia, this corpus study
addresses the question of the nature and the
structure of antonym and synonymy in language
use. While quite a lot of empirical research using
different observational techniques has been
carried on antonymy
The basic assumption underlying this study is that the strength of co-occurrence of antonyms and synonyms is dependent on the domain in which they are instantiated and co-occur. In order to test the hypothesis we mine the cooccurrence information of the antonyms and the synonyms relative to the domains using a dependency grammar method. 1 1 http://nlp.stanford.edu/software/lexparser.shtml The rationale is that the dependency parsing produces the relational information among the constituent words of a given sentence, which allows us to (i) extract co-occurrences specific to a given domain/context, and (ii) capture long distance co-occurrences between the word pairs. Consider (1).
1. Winters are cold and dry, summers are cool in the hills and quite hot in the plains. In (1), the antonyms cold: hot modify winters and summers respectively. Those forms express the lexical concepts winter and summer in the domain temperature. The antonyms cold: hot cooccur but at a distance in the sentence. Thanks to the dependency information, it is possible to extract such long distance co-occurrences together with the concepts modified.
The article is organized as follows. In section 2, we describe the procedure and the two methods used: co-occurrence extraction of lexical items in the same sentence and a variant domain dependent co-occurrence extraction method. The latter method extracts patterns of co-occurrence information of the synonyms and antonyms in different sentences. In section 3 we present the results and discussions followed by a discussion of our results in comparison with related previous works in section 4. The conclusions are presented in section 5. 2
Using an algorithm similar to the one proposed by Tesfaye & Zock (2012) and Zock & Tesfaye (2012), we extracted the co-occurrence information of the pairs in different domains separately, measuring the strength of their relation in the different domains with the aim of (i) making principled comparisons between antonyms and synonyms from a domain perspective, and (ii) determining the structure of antonymy and synonymy as categories in language and cognition.
Our algorithm is similar to the standard ngram co-occurrences extraction algorithms, but instead of using the linear ordering of the words in the text, it generates co-occurrences frequencies along paths in the dependency tree of the sentence as presented in the sections 2.2–2.5. 2.1
The antonyms and synonyms employed for
training and testing were extracted from the data used
by Paradis et al. (2009) where the antonyms are
presented according to their underlying
dimensions and synonyms were provided for all the
individual antonyms
Bad
The associated synonyms of the antonyms huge, vast, massive ,big ,bulky, giant ,gross, heavy, significant ,wide little, low, minor, minute, petite, slim, tiny quick, hurried, prompt, accelerating, rapid sudden, dull, gradual, lazy
muscular, powerful, substantial, tough Weak light, soft, thin, wimpy crappy, defective, evil ,harmful, poor ,shitty ,spoiled ,unhappy awful ,genuine ,great, honorable ,hot, neat, nice, reputable, right ,safe ,well
In order to extract the co-occurrences of the antonyms/synonyms in the respective domains we produced the relational information among the constituent words of a given sentence. To this end, we extracted the patterns linking the synonyms/antonyms and the concepts they modify and used this same pattern to extract more lexical concepts. The procedure was as follows.
Start with the selected set of synonym/antonym pairs Extract sentences containing the pairs Identify the dependency information of the sentences Mine the dependency patterns linking the pairs with the concepts they modify Use these learned patters to extract further relations (synonym/antonym pairs and the associated concepts) 2.3
We created a matrix of antonym and synonym pairs matching every antonym and synonym from the list in Table 1. Using the patterns learned in section 2.2 we identified as many domains as possible for the pairs of synonyms and antonyms and calculated their frequency of cooccurrence in the respective domains.
When the lexical concepts were considered too specific, we referred them to more inclusive, superordinate domains. Frequency of occurrence was used as a criterion for conflation of concepts into superordinate ones as follows.
Extract term co-occurrence frequencies within a window of sentences constituting both the antonyms/synonyms and the potential domain concepts. For instance: o Antonyms: cold: hot, domain
concepts: winter, summer o Synonyms: strong: heavy, do
main concepts: wind, rain Create a matrix of the potential domain concepts and the co-occurring terms with their frequencies Cluster them using the k-means algorithm Take the term with the maximal frequency (centroid) in each cluster and consider it the domain term Test the result using expert judgment running the algorithm on the test set. m y n o n y S / m y n o t n
A hot cold o D t p e ila con t c n n teo ia
P m summer win- temperature ter climate
Wind Words cooccurring with possible domain concepts 50 43 30 strong heavy wind rain wind rain 86 winds snow- winds snow- 3 fall fall winds rainfall winds rainfall 34 waves rain- waves rainfall 4 fall
The algorithm calculated the co-occurrence frequency of the antonyms/synonyms with the different concepts they refer to (or modify) as presented in table 3 by combining the information obtained in section 2.3 and section 2.4.
s m y n o t n
A
hot
cold
strong
heavy
In the previous algorithm, the co-occurrence
information was extracted from the same sentence.
However, unlike the antonyms, synonyms rarely
occurred together in the same context (the same
sentence and domain). It is natural to assume that
in most cases synonyms are used in different
contexts since they evoke similar but not
identical meanings. This is however not the case for
antonyms, which were always used to evoke
properties of the same meanings when these
antonymic words were used to express opposition
(Paradis & Willners 2011), and in fact also when
they are not used to express opposition
X and Z are a pair of a given antonym/synonym, Y is the domain within which the pairs of the antonym/synonym co-occur and f the frequency of the x-y or z-y co-occurrence.
The frequency of a pair of the antonyms/synonyms in the Y domain was counted and the same applies to the other pair. This made it possible to measure the degree of cooccurrence of the antonym/synonym pairs from the domain perspective indirectly.
3.1
Based on the results of the experiment the strength of the antonyms/synonyms varies in relation to the domains of instantiation. Hence, the strength of the co-occurrence of antonyms and synonyms is a function of the domains. For instance, the antonyms: slow: fast, slow: quick and slow: rapid were used in completely different domains with little or no overlap. Slow: fast is used in the domains of motion, movement, speed; slow: quick is used for time, march, steps domains. The synonyms powerful: strong are used in the domains of voices, links, meaning; strong: muscular in the domains of legs, neck; strong: heavy are used in the domains of wind rain, waves rainfall, winds snow respectively; intense: strong in the domains of battle resistance, radiation gravity, updrafts clouds respectively.
We observed some unique patterns among the antonyms and synonyms as described below: The antonyms: Co-occurred frequently in the same domain in the same sentence. The strength of the co-occurrence depends on the domain: slow: fast in the domains of growth, lines , motion, movement, speed ,trains, music, pitch; slow: quick in the domains of time, march, steps; slow: gradual in the domains of process, change, transition; small: big in the domains of screen, band; small: large in the domains of intestine, companies, businesses; week: strong in the domains of force, interaction, team, ties, points, sides, wind.
The Synonyms: Co-occurred in the same sentence but mainly in different domains. For instance, fast: quick, strong: heavy. Few co-occurrences in the same sentences in the same domains as exhibited by the pairs gradual: slow in the domains of process, change, development. The strength of the synonym cooccurrence depends on the domains. For instance, the synonyms strong: heavy in wind and rain domains respectively to express intensity; the synonyms large: wide in the domains of population and distribution domains respectively; gradual: slow in the domains of process, change, development; small: low in the domain of size cost, range, size weight, area, size price, amount density; micro: small in the domains of enterprises, businesses, entrepreneurs.. 3.2
As mentioned before, the variant domain dependent co-occurrence extraction algorithm mines the patterns of co-occurrence information of the synonyms and antonyms in different sentences. The result from the variant co-occurrence experiment showed hardly any differences in the domains with which the synonyms and antonyms are associated. Strong in the domains of influence, force, wind, interactions, evidence, ties; Heavy in the domains of loss, rain, industry, traffic; gradual: slow in the domains of process, change, transition. However, we observed that the frequency of co-occurrence differed significantly. For instance, the frequency of the pair gradual: slow was 76 in same sentences experiment but 1436 in the variant co-occurrence experiment.
Previous research has shown that there are
antonyms that are strongly opposing (canonical
antonyms)
Measuring the strength of antonyms without taking domains into account provided higher values for the canonicals as they tended to be used in several domains. If domains were taken in to account, as we did in this experiment, all the antonyms were strong in their specific domains. The antonym pair small: large had higher value without considering domain in to account yet had 0.29 value in the domain of screen where small: big has much higher value (0.71). The values were calculated taking the frequency of co-occurrence of the domain term (screen in this case) with each antonyms and dividing it by the summation of the frequency of co-occurrence of the domain term (again screen in this case) with both antonyms (small big and small large). 5
The strength of the antonyms/synonyms varied in relation to the domains of instantiation. The use of antonyms and synonyms was very consistent with few overlaps across the domains. Similar results were observed in both experiments from the domain perspective although with significant differences in frequency. Antonyms frequently co-occurred in the same domains in the same sentences and synonyms co-occurred in different domains in the same sentences (with less frequency) and more frequently in different sentences in the same domains.
We acknowledge European Science Foundation (ESF) for providing us the funding to undertake this work.