Finding Short Definitions of Terms on Web Pages

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1270–1279, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Finding Short Definitions of Terms on Web Pages Gerasimos Lampouras∗ and Ion Androutsopoulos∗+ ∗Department of Informatics, Athens University of Economics and Business, Greece +Digital Curation Unit, Research Centre “Athena”, Athens, Greece Abstract We present a system that finds short def- initions of terms on Web pages. It em- ploys a Maximum Entropy classifier, but it is trained on automatically generated ex- amples; hence, it is in effect unsupervised. We use ROUGE-W to generate training ex- amples from encyclopedias and Web snip- pets, a method that outperforms an alter- native centroid-based one. After training, our system can be used to find definitions of terms that are not covered by encyclo- pedias. The system outperforms a compa- rable publicly available system, as well as a previously published form of our system. 1 Introduction Definitions of terms are among the most com- mon types of information users search for on the Web. In the TREC 2001 QA track (Voorhees, 2001), where the distribution of question types re- flected that of real user logs, 27% of the ques- tions were requests for definitions (e.g., “What is gasohol?”, “Who was Duke Ellington?”). Conse- quently, some Web search engines provide special facilities (e.g., Google’s “define:” query prefix) that seek definitions of user-specified terms in on- line encyclopedias or glossaries; to save space, we call both “encyclopedias”. There are, however, of- ten terms that are too recent, too old, or less widely used to be included in encyclopedias. Their defi- nitions may be present on other Web pages (e.g., newspaper articles), but they may be provided in- directly (e.g., “He said that gasohol, a mixture of gasoline and ethanol, has been great for his busi- ness.”) and they may be difficult to locate with generic search engines that may return dozens of pages containing, but not defining the terms. We present a system to find short definitions of user-specified terms on Web pages. It can be used as an add-on to generic search engines, when no definitions can be found in on-line encyclope- dias. The system first invokes a search engine us- ing the (possibly multi-word) term whose defini- tion is sought, the target term, as the query. It then scans the top pages returned by the search engine to locate 250-character snippets with the target term at their centers; we call these snippets windows. The windows are candidate definitions of the target term, and they are then classified as acceptable (positive class) or unacceptable (nega- tive class) using supervised machine learning. The system reports the windows for which it is most confident that they belong in the positive class. Ta- ble 1 shows examples of short definitions found by our system. In our experiments, we allow the sys- tem to return up to five windows per target term, and the system’s response is counted as correct if any of the returned windows contains an …

Content truncated to 500 words