Open-domain Factoid Question Answering via Knowledge Graph Search

Proceedings of 2016 NAACL Human-Computer Question Answering Workshop, pages 22–28, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Open-domain Factoid Question Answering via Knowledge Graph Search Ahmad Aghaebrahimian, Filip Jurˇc´ıˇcek Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, 11800 Praha 1, Czech Republic {Ebrahimian, Jurcicek}@ufal.mff.cuni.cz Abstract We introduce a highly scalable approach for open-domain question answering with no de- pendence on any data set for surface form to logical form mapping or any linguistic ana- lytic tool such as POS tagger or named en- tity recognizer. We define our approach under the Constrained Conditional Models frame- work which lets us scale up to a full knowl- edge graph with no limitation on the size. On a standard benchmark, we obtained near 4 per- cent improvement over the state-of-the-art in open-domain question answering task. 1 Introduction We consider the task of simple open-domain ques- tion answering. The answer to a simple question can be obtained only by knowing one entity and one property (Bordes et al., 2015). A property is an attribute which is asked about a specific thing, place or people in a question. The thing, place or people in the question are samples of an entity. The answer to such a question is again an entity or a set of entities. For instance, in the question “What is the time zone in Dublin?”, Dublin is an entity and time zone is a property. Freebase (Bollacker et al., 2008), the knowledge graph which we used in our experiments, contains about 58 million such entities and more than 14 thousands such properties. Hence, entities which we obtain from the knowledge graph are ambiguous in majority of cases. We extract metadata available in the knowledge graph and integrate them into our system using Constrained Conditional Model frame- work (CCM) (Roth and Yih, 2005) to disambiguate entities. In WebQuestions (Berant et al., 2013), a data set of 5810 questions which are compiled using the Google suggest API, 86% of the questions are an- swerable by knowing only one entity (Bordes et al., 2015). It suggests that a large number of the ques- tions which ordinary people ask on the Internet are simple questions and it emphasizes on the impor- tance of simple question answering systems. Be- sides, the best result on this task is 63.9% (Bordes et al., 2015) which shows open-domain simple QA is still an unresolved task in NLP. Despite the title, simple QA is not a simple task at all. Flexible and unbound number of entities and their properties in open-domain questions is an in- timidating challenge for entity recognition. How- ever, knowledge graphs by providing a structural knowledge base on entities can help a lot. We use a knowledge graph to recognize entities at test time. Defining a model for entity disambigua- tion on a single question instead of a whole data set lets us to scale the system up to a large knowledge graph irrespective …

Content truncated to 500 words