A Constituent-Centric Neural Architecture for Reading Comprehension

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1405–1414 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1129 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1405–1414 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1129 A Constituent-Centric Neural Architecture for Reading Comprehension Pengtao Xie*† and Eric P. Xing† *Machine Learning Department, Carnegie Mellon University †Petuum Inc. pengtaox@cs.cmu.edu, eric.xing@petuum.com Abstract Reading comprehension (RC), aiming to understand natural texts and answer ques- tions therein, is a challenging task. In this paper, we study the RC problem on the Stanford Question Answering Dataset (SQuAD). Observing from the training set that most correct answers are centered around constituents in the parse tree, we design a constituent-centric neural archi- tecture where the generation of candidate answers and their representation learning are both based on constituents and guided by the parse tree. Under this architec- ture, the search space of candidate answers can be greatly reduced without sacrificing the coverage of correct answers and the syntactic, hierarchical and compositional structure among constituents can be well captured, which contributes to better rep- resentation learning of the candidate an- swers. On SQuAD, our method achieves the state of the art performance and the ab- lation study corroborates the effectiveness of individual modules. 1 Introduction Reading comprehension (RC) aims to answer questions by understanding texts, which is a chal- lenge task in natural language processing. Var- ious RC tasks and datasets have been devel- oped, including Machine Comprehension Test (Richardson et al., 2013) for multiple-choice ques- tion answering (QA) (Sachan et al., 2015; Wang and McAllester, 2015), Algebra (Hosseini et al., 2014) and Science (Clark and Etzioni, 2016) for passing standardized tests (Clark et al., 2016), CNN/Daily Mail (Hermann et al., 2015) and Chil- dren’s Book Test (Hill et al., 2015) for cloze-style The most authoritative account at the time came from the medical faculty in Paris in a report to the king of France that blamed the heavens. This report became the first and most widely circulated of a series of plague tracts that sought to give advice to sufferers. That the plague was caused by bad air became the most widely accepted theory. Today, this is known as the Miasma theory. 1. Who was the medical report written for? the king of France 2. What is the newer, more widely accepted theory behind the spread of the plague? bad air 3. What is the bad air theory officially known as? Miasma theory Figure 1: An example of the SQuAD QA task QA (Chen et al., 2016; Shen et al., 2016), Wik- iQA (Yang et al., 2015), Stanford Question An- swering Dataset (SQuAD) (Rajpurkar et al., 2016) and Microsoft Machine Reading Comprehension (Nguyen et al., 2016) for open domain QA. In this paper, we are specifically interested in solving the SQuAD QA task (Figure 1 shows an example), in light of its following features: (1) large scale: …

Content truncated to 500 words