Search-based Neural Structured Learning for Sequential Question Answering

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1821–1831 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1167 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1821–1831 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1167 Search-based Neural Structured Learning for Sequential Question Answering Mohit Iyyer∗ Department of Computer Science and UMIACS University of Maryland, College Park miyyer@umd.edu Wen-tau Yih, Ming-Wei Chang Microsoft Research Redmond, WA 98052 {scottyih,minchang}@microsoft.com Abstract Recent work in semantic parsing for ques- tion answering has focused on long and complicated questions, many of which would seem unnatural if asked in a normal conversation between two humans. In an ef- fort to explore a conversational QA setting, we present a more realistic task: answer- ing sequences of simple but inter-related questions. We collect a dataset of 6,066 question sequences that inquire about semi- structured tables from Wikipedia, with 17,553 question-answer pairs in total. To solve this sequential question answering task, we propose a novel dynamic neural semantic parsing framework trained using a weakly supervised reward-guided search. Our model effectively leverages the sequen- tial context to outperform state-of-the-art QA systems that are designed to answer highly complex questions. 1 Introduction Semantic parsing, which maps natural language text to meaning representations in formal logic, has emerged as a key technical component for building question answering systems (Liang, 2016). Once a natural language question has been mapped to a for- mal query, its answer can be retrieved by executing the query on a back-end structured database. One of the main focuses of semantic parsing research is how to address compositionality in lan- guage, and complicated questions have been specif- ically targeted in the design of a recently-released QA dataset (Pasupat and Liang, 2015). Take for ex- ample the following question: “of those actresses who won a Tony after 1960, which one took the most amount of years after winning the Tony to ∗Work done during an internship at Microsoft Research win an Oscar?” The corresponding logical form is highly compositional; in order to answer it, many sub-questions must be implicitly answered in the process (e.g., “who won a Tony after 1960?”). While we agree that semantic parsers should be able to answer very complicated questions, in re- ality these questions are rarely issued by users.1 Because users can interact with a QA system re- peatedly, there is no need to assume a single-turn QA setting where the exact question intent has to be captured with just one complex question. The same intent can be more naturally expressed through a sequence of simpler questions, as shown below: 1. What actresses won a Tony after 1960? 2. Of those, who later won an Oscar? 3. Who had the biggest gap between their two award wins? Decomposing complicated intents into multiple re- lated but simpler questions is arguably a more ef- fective strategy to explore a topic of interest, and it reduces the …

Content truncated to 500 words