Question Answering with Lexical Chains Propagating Verb Arguments

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 897–904, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Question Answering with Lexical Chains Propagating Verb Arguments Adrian Novischi Dan Moldovan Language Computer Corp. 1701 N. Collins Blvd, Richardson, TX, 75080 � adrian,moldovan � @languagecomputer.com Abstract This paper describes an algorithm for propagating verb arguments along lexi- cal chains consisting of WordNet rela- tions. The algorithm creates verb argu- ment structures using VerbNet syntactic patterns. In order to increase the cover- age, a larger set of verb senses were auto- matically associated with the existing pat- terns from VerbNet. The algorithm is used in an in-house Question Answering sys- tem for re-ranking the set of candidate an- swers. Tests on factoid questions from TREC 2004 indicate that the algorithm im- proved the system performance by 2.4%. 1 Introduction In Question Answering the correct answer can be formulated with different but related words than the question. Connecting the words in the ques- tion with the words in the candidate answer is not enough to recognize the correct answer. For ex- ample the following question from TREC 2004 (Voorhees, 2004): Q: (boxer Floyd Patterson) Who did he beat to win the title? has the following wrong answer: WA: He saw Ingemar Johanson knock down Floyd Patterson seven times there in winning the heavyweight title. Although the above sentence contains the words Floyd, Patterson, win, title, and the verb beat can be connected to the verb knock down using lexical chains from WordNet, this sentence does not an- swer the question because the verb arguments are in the wrong position. The proposed answer de- scribes Floyd Patterson as being the object/patient of the beating event while in the question he is the subject/agent of the similar event. Therefore the selection of the correct answer from a list of candidate answers requires the check of additional constraints including the match of verb arguments. Previous approaches to answer ranking, used syntactic partial matching, syntactic and semantic relations and logic forms for selecting the correct answer from a set of candidate answers. Tanev et al. (Tanev et al., 2004) used an algorithm for partial matching of syntactic structures. For lexi- cal variations they used a dependency based the- saurus of similar words (Lin, 1998). Hang et al. (Cui et al., 2004) used an algorithm to compute the similarity between dependency relation paths from a parse tree to rank the candidate answers. In TREC 2005, Ahn et al. (Ahn et al., 2005) used Discourse Representation Structures (DRS) resembling logic forms and semantic relations to represent questions and answers and then com- puted a score “indicating how well DRSs match each other”. Moldovan and Rus (Moldovan and Rus, 2001) transformed the question and the can- didate answers into logic forms and used a logic prover to determine if the candidate answer logic form (ALF) entails the question logic form(QLF). Continuing this work Moldovan et al. (Moldovan et al., 2003) built a …

Content truncated to 500 words