SW-AG: Local Context Matching for English Lexical Substitution

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 304–307, Prague, June 2007. c⃝2007 Association for Computational Linguistics SW-AG: Local Context Matching for English Lexical Substitution George Dahl, Anne-Marie Frassica, Richard Wicentowski Department of Computer Science Swarthmore College Swarthmore, PA 19081 USA {george.dahl, afrassi1}@gmail.com, richardw@cs.swarthmore.edu Abstract We present two systems that pick the ten most appropriate substitutes for a marked word in a test sentence. The first system scores candidates based on how frequently their local contexts match that of the marked word. The second system, an enhancement to the first, incorporates cosine similarity us- ing unigram features. The core of both sys- tems bypasses intermediate sense selection. Our results show that a knowledge-light, di- rect method for scoring potential replace- ments is viable. 1 Introduction An obvious way to view the problem of lexical sub- stitution is as a sense disambiguation task. For example, one possible approach is to identify the sense of the target word and then to pick a synonym based on the identified sense. Following Dagan et al. (2006), we refer to this as an indirect approach. A system using an indirect approach must have ac- cess to a list of senses for each target word, and each sense must have a corresponding list of synonyms. Though one can use a predefined sense inventory, such as WordNet, the granularity of the sense inven- tory may not be appropriate for the task. If the sense inventory is too fine-grained, then picking the cor- rect sense may be needlessly difficult. Conversely, if it is too coarse, picking the correct sense may not narrow down the list of potential substitutions suffi- ciently. To avoid these problems, we propose a direct ap- proach, which will break the problem into two steps: for each target word, generate a list of candidate syn- onyms; then rank each synonym for its quality as a replacement. Although our second system makes use of some sense information, it is used only to re- rank candidates generated using a direct approach. We describe two systems: the first is a purely di- rect method based on local context matching, and the second is a hybrid of local context matching and wider context bag-of-words matching. Both are knowledge-light and unsupervised. 2 Methods As mentioned above, we divide the task into two steps: compiling a list of synonyms and then, for each test instance, ranking the list of appropriate synonyms. Both of our systems create lists of can- didate synonyms in the same way and only differ in the way they arrive at a ranking for these candidates. 2.1 Compiling a substitution lexicon We begin by compiling a list of candidate synonyms for each target word. Following Dagan et al. (2006), we will refer to this list of synonyms as our substitu- tion lexicon. The performance of our system is lim- ited by the substitution lexicon because it can only pick the correct replacements if they are in the lexi- con. The substitution lexicon available to our …

Content truncated to 500 words