Queries as a Source of Lexicalized Commonsense Knowledge

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1081–1091, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Queries as a Source of Lexicalized Commonsense Knowledge Marius Pas¸ca Google Inc. 1600 Amphitheatre Parkway Mountain View, California 94043 mars@google.com Abstract The role of Web search queries has been demonstrated in the extraction of attributes of instances and classes, or of sets of re- lated instances and their class labels. This paper explores the acquisition of open- domain commonsense knowledge, usu- ally available as factual knowledge, from Web search queries. Similarly to previ- ous work in open-domain information ex- traction, knowledge extracted from text - in this case, from queries - takes the form of lexicalized assertions associated with open-domain classes. Experimental results indicate that facts extracted from queries complement, and have competitive accuracy levels relative to, facts extracted from Web documents by previous meth- ods. 1 Introduction Motivation: Open-domain information extrac- tion methods (Etzioni et al., 2005; Pennac- chiotti and Pantel, 2009; Wang and Cohen, 2009; Kozareva and Hovy, 2010; Wu et al., 2012) aim at distilling text into knowledge assertions about classes, instances and relations among them (Et- zioni et al., 2011). Ideally, the assertions would complement or expand upon knowledge avail- able in popular, human-created resources such as Wikipedia (Remy, 2002) and Freebase (Bollacker et al., 2008), reducing costs and scalability is- sues associated with manual editing, curation and maintenance of knowledge. Candidate knowledge assertions extracted from text for various instances and classes (Banko et al., 2007; Cafarella et al., 2008; Wu and Weld, 2010) must satisfy several constraints in order to be use- ful. First, their boundaries must be correctly iden- tified within the larger context (e.g., a document sentence) from which they are extracted. In prac- tice, this is a challenge with arbitrary Web docu- ments, where even instances and class labels that are complex nouns, and thus still shorter than can- didate assertions, are difficult to precisely detect and pick out from surrounding text (Downey et al., 2007). This causes the extraction of assertions like companies may “be in the process”, hurri- canes may “run from june”, or video games may “make people” (Fader et al., 2011). Second, the assertions must be correctly associated with their corresponding instance or class. In practice, tag- ging and parsing errors over documents of arbi- trary quality may cause the extracted assertions to be associated with the wrong instances or classes. Examples are video games may “watch movies”, or video games may “read a book”. Third, the as- sertions, even if true, must refer to relevant prop- erties or facts, rather than to statements of little or no practical interest to anyone. In practice, relevant properties may be difficult to distinguish from uninteresting statements in Web documents. Consequently, assertions extracted from Web doc- uments include the facts that companies may “say in a statement”, or that hurricanes may “be just around the corner” or may “be in effect”. Contributions: This paper explores the …

Content truncated to 500 words