Answering Definition Questions via Temporally-Anchored Text Snippets

Answering Definition Questions via Temporally-Anchored Text Snippets Marius Pas¸ca Google Inc. 1600 Amphitheatre Parkway Mountain View, California 94043 mars@google.com Abstract A lightweight extraction method derives text snippets associated to dates from the Web. The snippets are organized dynamically into answers to definition questions. Experi- ments on standard test question sets show that temporally-anchored text snippets allow for efficiently answering definition ques- tions at accuracy levels comparable to the best systems, without any need for complex lexical resources, or specialized processing modules dedicated to finding definitions. 1 Introduction In the field of automated question answering (QA), a variety of information sources and multiple extrac- tion techniques can all contribute to producing rele- vant answers in response to natural-language ques- tions submitted by users. Yet the nature of the infor- mation source which is mined for answers, together with the scope of the questions, have the most sig- nificant impact on the overall architecture of a QA system. When compared to the average queries sub- mitted in a decentralized information seeking envi- ronment such as Web search, fact-seeking questions tend to specify better the nature of the information being sought by the user, whether it is the name of the longest river in some country, or the name of the general who defeated the Spanish Armada. In order to understand the structure and the linguistic clues encoded in natural-language questions, many QA systems employ sophisticated techniques, thus deriving useful information such as terms, relations among terms, the type of the expected answers (e.g., cities vs. countries vs. presidential candidates), and other semantic constraints (e.g., the elections from 1978 rather than any other year). One class of questions whose characteristics place them closer to exploratory queries, rather than stan- dard fact-seeking questions, are definition questions. Seeking information about an entity or a concept, questions such as “Who is Caetano Veloso?” of- fer little guidance as to what particular techniques could be used in order to return relevant information from a large text collection. In fact, the same user may choose to submit a definition question or a sim- pler exploratory query (Caetano Veloso), and still look for text snippets capturing relevant properties of the question concept. Various studies (Chen et al., 2006; Han et al., 2006) illustrate the challenges in- troduced by definition questions. As such questions have a less irregular form than other open-domain questions, recognizing their type is relatively eas- ier (Hildebrandt et al., 2004). Conversely, the iden- tification of relevant documents and the extraction of answers to definition questions are more labori- ous, and the impact on the architecture of QA sys- tems is quite significant. Indeed, separate, dedicated modules, or even end-to-end systems are specifi- cally built for answering definition questions (Kla- vans and Mures¸an, 2001; Hildebrandt et al., 2004; Greenwood and Saggion, 2004). The importance of definition questions among other question categories is confirmed by their inclusion among the evalua- tion queries from the QA track of TREC evalua- tions (Voorhees and Tice, 2000). This paper investigates the …

Content truncated to 500 words