Supervised Synonym Acquisition Using Distributional Features and Syntactic Patterns

Abstract

Distributional similarity has been widely used to capture the semantic relatedness of words in many NLP tasks. However, parameters such as similarity measures must be manually tuned to make distributional similarity work effectively. To address this problem, we propose a novel approach to synonym identification based on supervised learning and distributional features, which correspond to the commonality of individual context types shared by word pairs. This approach also enables the integration with pattern-based features. In our experiment, we have built and compared eight synonym classifiers, and showed a drastic performance increase of over 60% on F-1 measure, compared to the conventional similarity-based classification. Distributional features that we have proposed are better in classifying synonyms than the conventional common features, while the pattern-based features have appeared almost redundant.

Comparing Distributional and Mirror Translation Similarities for Extracting Synonyms

“…Our longer-term objective is to reproduce synonymy word pair supervised classification; any similarity alone scores quite low as a synonymy descriptor, but experiments, such as [17], show it is doable to reliably label word pairs with lexical functions if the proportion of candidates is more balanced than the very low natural proportion, and this means designing a filter as we do here.…”

“…A different setting is resampled evaluation, where a classifier is trained and tested on a set of word pairs with an priori ratio of synonyms and non-synonyms. It is only relevant if a good preselection method allows one to reach the assumed proportions of synonyms in the training and test sets [17]. Our results could actually be considered as an input to such methods.…”

SurfCon

“…Synonym Discovery. A variety of methods have been proposed to detect synonyms of medical terms, ranging from utilizing lexical patterns [39] and clustering [21] to the distributional semantics models [11]. There are some more recent works on automatic synonym discovery [30,31,37,42].…”

Early and Late Combinations of Criteria for Reranking Distributional Thesauri

“…However, relying only on these two criteria doesn't lead to a good enough set of positive examples. For instance, taking as positive examples from the initial thesaurus of Section 4 the first neighbor of its 2,148 most frequent entries, the number of positive examples of (Hagiwara et al, 2009), only leads to 44.3% of correct examples. Moreover, this percentage exceeds 50% only when the number of examples is less than 654, which represents a very small training set for this kind of task.…”