Machine Comprehension with Discourse Relations

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1253–1262, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Machine Comprehension with Discourse Relations Karthik Narasimhan CSAIL, MIT karthikn@csail.mit.edu Regina Barzilay CSAIL, MIT regina@csail.mit.edu Abstract This paper proposes a novel approach for incorporating discourse information into machine comprehension applications. Traditionally, such information is com- puted using off-the-shelf discourse analyz- ers. This design provides limited oppor- tunities for guiding the discourse parser based on the requirements of the target task. In contrast, our model induces re- lations between sentences while optimiz- ing a task-specific objective. This ap- proach enables the model to benefit from discourse information without relying on explicit annotations of discourse structure during training. The model jointly iden- tifies relevant sentences, establishes rela- tions between them and predicts an an- swer. We implement this idea in a discrim- inative framework with hidden variables that capture relevant sentences and rela- tions unobserved during training. Our ex- periments demonstrate that the discourse aware model outperforms state-of-the-art machine comprehension systems.1 1 Introduction The task of machine comprehension concerns the automatic extraction of answers from a given pas- sage. Often, the relevant information required to answer a question is distributed across multiple sentences. Understanding the relation(s) between these sentences is key to finding the correct an- swer. Consider the example in fig. 1. To answer the question about why Sally put on her shoes , we need to infer that She put on her shoes and She went outside to walk are connected by a causality relation. 1Code and data are available at http://people. csail.mit.edu/karthikn/mcdr. Sally liked going outside. She put on her shoes. She went outside to walk. [...] Missy the cat meowed to Sally. Sally waved to Missy the cat. [...] Sally hears her name. ”Sally, Sally, come home”, Sally’s mom calls out. Sally runs home to her Mom. Sally liked going outside. Why did Sally put on her shoes? A) To wave to Missy the cat B) To hear her name C) Because she wanted to go outside D) To come home Figure 1: Sample story excerpt from a passage in the MCTest dataset.2 Correct answer is in italics. Prior work has demonstrated the value of dis- course relations in related applications such as question answering (Jansen et al., 2014). Tradi- tionally, however, these approaches rely on out- puts from off-the-shelf discourse analyzers, us- ing them as features for target applications. Such pipeline designs provide limited opportunities for guiding the discourse parser based on the require- ments of the end task. Given a wide spectrum of discourse frameworks (Mann and Thompson, 1988; Prasad et al., 2008; Wolf and Gibson, 2005), it is not clear a priori what the optimal set of dis- course annotations is for the task. Moreover, a generic discourse parser may introduce additional errors due to the mismatch between its training corpus and a dataset used in an application. In fact, the …

Content truncated to 500 words