Learning to Answer Questions from Wikipedia Infoboxes

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1930–1935, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Learning to Answer Questions from Wikipedia Infoboxes Alvaro Morales CSAIL MIT alvarom@mit.edu Varot Premtoon CSAIL MIT varot@mit.edu Cordelia Avery CSAIL MIT cavery@mit.edu Sue Felshin CSAIL MIT sfelshin@mit.edu Boris Katz CSAIL MIT boris@mit.edu Abstract A natural language interface to answers on the Web can help us access information more ef- ficiently. We start with an interesting source of information—infoboxes in Wikipedia that summarize factoid knowledge—and develop a comprehensive approach to answering ques- tions with high precision. We first build a system to access data in infoboxes in a struc- tured manner. We use our system to construct a crowdsourced dataset of over 15,000 high- quality, diverse questions. With these ques- tions, we train a convolutional neural network model that outperforms models that achieve top results in similar answer selection tasks. 1 Introduction The goal of open-domain question answering is to provide high-precision access to information. With many sources of knowledge on the Web, selecting the right answer to a user’s question remains chal- lenging. Wikipedia contains over five million arti- cles in its English version. Providing a natural lan- guage interface to answers in Wikipedia is an impor- tant step towards more effective information access. Many Wikipedia articles have an infobox, a table that summarizes key information in the article in the form of attribute–value pairs like “Narrated by: Fred Astaire”. This data source is appealing for question answering because it covers a broad range of facts that are inherently relevant: a human editor manually highlighted this information in the infobox. Although many infoboxes appear to be similar, they are only semi-structured—few attributes have Q Who took over after Nelson Mandela? A Succeeded by: Thabo Mbeki Q Who designed Central Park? A Architect: Frederick Law Olmsted Q Where did Oscar Wilde earn his degree? A Alma mater: Trinity College, Dublin Q What does Intel do? A Industry: Semiconductors Table 1: Example questions and answers with little lexical overlap from the INFOBOXQA dataset. consistent value types across articles, infobox tem- plates do not mandate which attributes must be included, and editors are allowed to add article- specific attributes. Infobox-like tables are very com- mon on the Web. Since it is infeasible to incorporate every such source into structured knowledge bases like Freebase (Bollacker et al., 2008), we need tech- niques that do not rely on ontology or value type information. We focus on the answer selection problem, where the goal is to select the best answer out of a given candidate set of attribute–value pairs from infoboxes corresponding to a named entity in the question. Ta- ble 1 illustrates how questions from users may have little lexical overlap with the correct attribute–value pair. Answer selection is an important subtask in building an end-to-end question answering system. Our work has two main contributions: (1) We compiled the INFOBOXQA dataset, a crowdsourced corpus of over 15,000 questions with answers from …

Content truncated to 500 words