Open Information Extraction from Question-Answer Pairs

Proceedings of NAACL-HLT 2019, pages 2294–2305 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2294 Open Information Extraction from Question-Answer Pairs Nikita Bhutani ∗ University of Michigan nbhutani@umich.edu Yoshihiko Suhara Megagon Labs yoshi@megagon.ai Wang-Chiew Tan Megagon Labs wangchiew@megagon.ai Alon Halevy Megagon Labs alon@megagon.ai H. V. Jagadish University of Michigan jag@eecs.umich.edu Abstract Open Information Extraction (OPENIE) ex- tracts meaningful structured tuples from free- form text. Most previous work on OPENIE considers extracting data from one sentence at a time. We describe NEURON, a system for extracting tuples from question-answer pairs. Since real questions and answers often con- tain precisely the information that users care about, such information is particularly desir- able to extend a knowledge base with. NEURON addresses several challenges. First, an answer text is often hard to understand without knowing the question, and second, rel- evant information can span multiple sentences. To address these, NEURON formulates extrac- tion as a multi-source sequence-to-sequence learning task, wherein it combines distributed representations of a question and an answer to generate knowledge facts. We describe exper- iments on two real-world datasets that demon- strate that NEURON can find a significant num- ber of new and interesting facts to extend a knowledge base compared to state-of-the-art OPENIE methods. 1 Introduction Open Information Extraction (OPENIE) (Banko et al., 2007) is the problem of extracting structured data from a text corpus, without knowing a priori which relations will be extracted. It is one of the primary technologies used in building knowledge bases (KBs) that, in turn, power question answer- ing (Berant et al., 2013). The vast majority of pre- vious work on OPENIE extracts structured informa- tion (e.g., triples) from individual sentences. This paper addresses the problem of extract- ing structured data from conversational question- answer (CQA) data. Often, CQA data contains pre- cisely the knowledge that users care about. As ∗ Part of the work was done while the author was at Megagon Labs. such, this data offers a goal-directed method for extending existing knowledge bases. Consider, for example, a KB about a hotel that is used to power its website and/or a conversational interface for hotel guests. The KB provides information about the hotel’s services: complimentary breakfast, free wifi, spa. However, it may not include information about the menu/times for the breakfast, credentials for the wifi, or the cancellation policy for a spa appointment at the hotel. Given the wide range of information that may be of interest to guests, it is not clear how to extend the KB in the most effective way. However, the conversational logs, which many hotels keep, contain the actual ques- tions from guests, and can therefore be used as a resource for extending the KB. Following exam- ples illustrate the kind of data we aim to extract: Example 1. Q: Does the hotel have a gym? A: It is located on the third floor and is 24/7. Tuple: ⟨gym, is located on, third floor⟩ Example 2. Q: What time does the pool open? …

Content truncated to 500 words