Simple Question Answering by Attentive Convolutional Neural Network

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1746–1756, Osaka, Japan, December 11-17 2016. Simple Question Answering by Attentive Convolutional Neural Network∗ Wenpeng Yin∗, Mo Yu†, Bing Xiang†, ∗Center for Information and Language Processing LMU Munich, Germany wenpeng@cis.lmu.de Bowen Zhou†, Hinrich Sch¨utze∗ †IBM Watson Yorktown Heights, NY, USA {yum,bingxia,zhou}@us.ibm.com Abstract This work focuses on answering single-relation factoid questions over Freebase. Each question can acquire the answer from a single fact of form (subject, predicate, object) in Freebase. This task, simple question answering (SimpleQA), can be addressed via a two-step pipeline: entity linking and fact selection. In fact selection, we match the subject entity in a fact candidate with the entity mention in the question by a character-level convolutional neural network (char-CNN), and match the predicate in that fact with the question by a word-level CNN (word-CNN). This work makes two main contributions. (i) A simple and effective entity linker over Freebase is proposed. Our entity linker outperforms the state-of-the-art entity linker over SimpleQA task. 1 (ii) A novel attentive maxpooling is stacked over word-CNN, so that the predicate representation can be matched with the predicate-focused question representation more effectively. Experiments show that our system sets new state-of-the-art in this task. 1 Introduction Factoid question answering (QA) over knowledge bases such as Freebase (Bollacker et al., 2008) has been intensively studied recently (e.g., Bordes et al. (2014), Yao et al. (2014), Bast and Haussmann (2015), Yih et al. (2015), Xu et al. (2016)). Answering a question can require reference to multiple related facts in Freebase or reference to a single fact. This work studies simple question answering (SimpleQA) based on the SimpleQuestions benchmark (Bordes et al., 2015) in which answering a question does not require reasoning over multiple facts. Single-relation factual questions are the most common type of question observed in various community QA sites (Fader et al., 2013) and in search query logs. Even though this task is called “simple”, it is in reality not simple at all and far from solved. In SimpleQA, a question, such as “what’s the hometown of Obama?”, asks a single and direct topic of an entity. In this example, the entity is “Obama” and the topic is hometown. So our task is reduced to finding one fact (subject, predicate, object) in Freebase that answers the question, which roughly means the subject and predicate are the best matches for the topical entity “Obama” and for the topic description “what’s the hometown of”, respectively. Thus, we aim to design a method that picks a fact from Freebase, so that this fact matches the question best. This procedure resembles answer selection (Yu et al., 2014) in which a system, given a question, is asked to choose the best answer from a list of candidates. In this work, we formulate the SimpleQA task as a fact selection problem and the key issue lies in the system design for how to match a fact candidate to the question. The first obstacle is that …

Content truncated to 500 words