Learning Lexical Semantic Relations using Lexical Analogies — Extended Abstract

Linguistic ontologies, most notably WordNet [1], have been shown to be a valuable resource for a variety of natural language processing applications. Presently, linguistic ontologies are largely constructed by hand, which is both difficult and expensive. A central problem that demands an automated solution is the discovery and incorporation of lexical semantic relations, or semantic relations between concepts. Lexical semantic relations are the fundamental building blocks that allow words to be associated with each other and linked together to form cohesive text. Despite their importance, lexical semantic relations are severely underrepresented in current linguistic ontologies. As Morris and Hirst [2] point out, current linguistic ontologies only capture what they call classical relations — basically, WordNet relations such as hyponymy, hypernymy, troponymy, meronymy, antonymy, and synonymy. However, the majority of lexical semantic relations found in real-world text are in fact non-classical — for example, positive-qualities (humbleness and kindness), cause-of (alcohol and drunk), and founder-of (Gate and Microsoft) [2]. Manually populating linguistic ontologies with all instances of classical and non-classical relations is impractical as there are simply too many of them and there are as yet no systematic methods for even recognizing non-classical relations in an arbitrary text. Clearly, automation is needed. In this research, we tackle the problem of automated learning of lexical semantic relations from text. We present an iterative algorithm in Sect. 2 that expands a small set of sample relation instances to a much larger set by making use of a dictionary of lexical analogies. In Sect. 3, we demonstrate a system that generates this dictionary automatically from text. The system builds lexical analogies by computing the similarities of the semantic relations between words, which we characterize by their dependency structures. The actual computation of similarity is carried out using a Vector Space Model augmented with Singular Value Decomposition. We give some promising preliminary experimental results in Sect. 4, and conclude with an outline of future work in Sect. 5.