Exploiting Strong Syntactic Heuristics and Co-Training to Learn Semantic Lexicons

Exploiting Strong Syntactic Heuristics and Co-Training to Learn Semantic Lexicons William Phillips and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 USA {phillips,riloff}@cs.utah.edu Abstract We present a bootstrapping method that uses strong syntactic heuristics to learn semantic lexicons. The three sources of information are appositives, compound nouns, and ISA clauses. We apply heuris- tics to these syntactic structures, embed them in a bootstrapping architecture, and combine them with co-training. Results on WSJ articles and a pharmaceutical cor- pus show that this method obtains high precision and finds a large number of terms. 1 Introduction Syntactic structure helps us understand the seman- tic relationships between words. Given a text cor- pus, we can use knowledge about syntactic struc- tures to obtain semantic knowledge. For example, Hearst (Hearst, 1992) learned hyponymy relation- ships by collecting words in lexico-syntactic expres- sions, such as “NP, NP, and other NPs”, and Roark and Charniak (Roark and Charniak, 1998) gener- ated semantically related words by applying statisti- cal measures to syntactic contexts involving apposi- tives, lists, and conjunctions. Exploiting syntactic structures to learn semantic knowledge holds great promise, but can run into problems. First, lexico-syntactic expressions that explicitly indicate semantic relationships (e.g., “NP, NP, and other NPs”) are reliable but a lot of semantic information occurs outside these expressions. Sec- ond, general syntactic structures (e.g., lists and con- junctions) capture a wide range of semantic rela- tionships. For example, conjunctions frequently join items of the same semantic class (e.g., “cats and dogs”), but they can also join different seman- tic classes (e.g., “fire and ice”). Some researchers (Roark and Charniak, 1998; Riloff and Shepherd, 1997) have applied statistical methods to identify the strongest semantic associations. This approach has produced reasonable results, but the accuracy of these techniques still leaves much room for improve- ment. We adopt an intermediate approach that learns semantic lexicons using strong syntactic heuristics, which are both common and reliable. We have identified certain types of appositives, compound nouns, and identity (ISA) clauses that indicate spe- cific semantic associations between words. We em- bed syntactic heuristics in a bootstrapping process and present empirical results demonstrating that this bootstrapping process produces high-quality seman- tic lexicons. In another set of experiments, we in- corporate a co-training (Blum and Mitchell, 1998) mechanism to combine the hypotheses generated by different types of syntactic structures. Co-training produces a synergistic effect across different heuris- tics, substantially increasing the coverage of the lex- icons while maintaining nearly the same level of ac- curacy. 2 Semantic Lexicon Learning The goal of our research is to automatically gener- ate a semantic lexicon. For our purposes, we de- fine a semantic lexicon to be a list of words with semantic category labels. For example, the word “bird” might be labeled as an ANIMAL and the word “car” might be labeled as a VEHICLE. Semantic lexicons have proven to be useful for many lan- Association for Computational Linguistics. Language Processing (EMNLP), Philadelphia, July 2002, pp. 125-132. Proceedings of the …

Content truncated to 500 words