Improving Machine Reading Comprehension with Contextualized Commonsense Knowledge

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8736 - 8747 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Improving Machine Reading Comprehension with Contextualized Commonsense Knowledge Kai Sun1† Dian Yu2† Jianshu Chen2 Dong Yu2 Claire Cardie1 1Cornell University, Ithaca, NY 2Tencent AI Lab, Bellevue, WA ks985@cornell.edu, {yudian, jianshuchen, dyu}@tencent.com, cardie@cs.cornell.edu Abstract To perform well on a machine reading compre- hension (MRC) task, machine readers usually require commonsense knowledge that is not explicitly mentioned in the given documents. This paper aims to extract a new kind of struc- tured knowledge from scripts and use it to im- prove MRC. We focus on scripts as they con- tain rich verbal and nonverbal messages, and two relevant messages originally conveyed by different modalities during a short time period may serve as arguments of a piece of com- monsense knowledge as they function together in daily communications. To save human ef- forts to name relations, we propose to repre- sent relations implicitly by situating such an argument pair in a context and call it contextu- alized knowledge. To use the extracted knowledge to improve MRC, we compare several fine-tuning strate- gies to use the weakly-labeled MRC data con- structed based on contextualized knowledge and further design a teacher-student paradigm with multiple teachers to facilitate the trans- fer of knowledge in weakly-labeled MRC data. Experimental results show that our paradigm outperforms other methods that use weakly- labeled data and improves a state-of-the-art baseline by 4.3% in accuracy on a Chinese multiple-choice MRC dataset C3, wherein most of the questions require unstated prior knowledge. We also seek to transfer the knowledge to other tasks by simply adapting the resulting student reader, yielding a 2.9% improvement in F1 on a relation extraction dataset DialogRE, demonstrating the poten- tial usefulness of the knowledge for non-MRC tasks that require document comprehension. 1 Introduction Given a question and a document, machine read- ing comprehension (MRC) tasks usually aim to † Work was conducted when K. S. was an intern at the Tencent AI Lab, Bellevue, WA. Equal contribution. select the correct answer option(s) from all op- tions (Richardson et al., 2013; Lai et al., 2017) or extract a span as the answer from the docu- ment (Hermann et al., 2015; Rajpurkar et al., 2016). For many MRC tasks, which are mostly in multiple- choice formats, machine readers require extensive general world knowledge unstated in the given doc- uments to perform well on tasks. Recent stud- ies have showed the usefulness of other human- annotated MRC datasets (Chung et al., 2018) or data from relevant tasks such as natural language inference (Yin et al., 2021). There is also a trend towards taking advantage of existing crowdsourced general world knowledge graphs such as Con- ceptNet (Speer et al., 2017) or automatically con- structed graphs (Zhang et al., 2020) to improve MRC tasks (Mostafazadeh et al., 2016; Ostermann et al., 2018; Sun et al., 2019a; Huang et al., 2019). However, most progress has …

Content truncated to 500 words