CFO: Conditional Focused Neural Question Answering with Large-scale Knowledge Bases

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 800–810, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics CFO: Conditional Focused Neural Question Answering with Large-scale Knowledge Bases Zihang Dai∗ Carnegie Mellon University dzihang@andrew.cmu.edu Lei Li∗ Toutiao.com lileicc@gmail.com Wei Xu Baidu Research xuwei06@baidu.com Abstract How can we enable computers to automat- ically answer questions like “Who created the character Harry Potter”? Carefully built knowledge bases provide rich sources of facts. However, it remains a chal- lenge to answer factoid questions raised in natural language due to numerous ex- pressions of one question. In particular, we focus on the most common questions — ones that can be answered with a sin- gle fact in the knowledge base. We pro- pose CFO, a Conditional Focused neural- network-based approach to answering fac- toid questions with knowledge bases. Our approach first zooms in a question to find more probable candidate subject men- tions, and infers the final answers with a unified conditional probabilistic frame- work. Powered by deep recurrent neural networks and neural embeddings, our pro- posed CFO achieves an accuracy of 75.7% on a dataset of 108k questions – the largest public one to date. It outperforms the cur- rent state of the art by an absolute margin of 11.8%. 1 Introduction Community-driven question answering (QA) web- sites such as Quora, Yahoo-Answers, and An- swers.com are accumulating millions of users and hundreds of millions of questions. A large portion of the questions are about facts or trivia. It has been a long pursuit to enable machines to answer such questions automatically. In recent years, several efforts have been made on utilizing open-domain knowledge bases to answer factoid questions. A knowledge ∗Part of the work was done while at Baidu. base (KB) consists of structured representation of facts in the form of subject-relation-object triples. Lately, several large-scale general- purpose KBs have been constructed, including YAGO (Suchanek et al., 2007), Freebase (Bol- lacker et al., 2008), NELL (Carlson et al., 2010), and DBpedia (Lehmann et al., 2014). Typically, structured queries with predefined semantics (e.g. SPARQL) can be issued to retrieve specified facts from such KBs. Thus, answering factoid questions will be straightforward once they are converted into the corresponding structured form. However, due to complexity of language, converting natural language questions to structure forms remains an open challenge. Among all sorts of questions, there is one cat- egory that only requires a single fact (triple) in KB as the supporting evidence. As a typical ex- ample, the question “Who created the charac- ter Harry Potter” can be answered with the sin- gle fact (HarryPotter, CharacterCreatedBy, J.K.Rowling). In this work, we refer to such questions as single-fact questions. Previously, it has been observed that single-fact questions con- stitute the majority of factoid questions in commu- nity QA sites (Fader et al., 2013). Despite the sim- plicity, automatically answering such questions re- mains far from solved — the latest best result on a dataset of 108k single-fact questions is only 63.9% …

Content truncated to 500 words