Lite Unified Modeling for Discriminative Reading Comprehension

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8682 - 8695 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Lite Unified Modeling for Discriminative Reading Comprehension Yilin Zhao1,2, Hai Zhao1,2,∗, Libin Shen3, Yinggong Zhao3 1 Department of Computer Science and Engineering, Shanghai Jiao Tong University 2 Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, China 3Leyan Tech, Shanghai, China zhaoyilin@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn libin@leyantech.com, ygzhao@leyantech.com Abstract As a broad and major category in machine reading comprehension (MRC), the generalized goal of discriminative MRC is answer predic- tion from the given materials. However, the focuses of various discriminative MRC tasks may be diverse enough: multi-choice MRC requires model to highlight and integrate all potential critical evidence globally; while ex- tractive MRC focuses on higher local bound- ary preciseness for answer extraction. Among previous works, there lacks a unified design with pertinence for the overall discriminative MRC tasks. To fill in above gap, we pro- pose a lightweight POS-Enhanced Iterative Co-Attention Network (POI-Net) as the first attempt of unified modeling with pertinence, to handle diverse discriminative MRC tasks synchronously. Nearly without introducing more parameters, our lite unified design brings model significant improvement with both en- coder and decoder components. The evalua- tion results on four discriminative MRC bench- marks consistently indicate the general effec- tiveness and applicability of our model, and the code is available at https://github. com/Yilin1111/poi-net. 1 Introduction Machine reading comprehension (MRC) as a chal- lenging branch in NLU, has two major categories: generative MRC which emphasizes on answer gen- eration (Koˇciský et al., 2018), and discriminative MRC which focuses on answer prediction from given contexts (Baradaran et al., 2020). Among them, discriminative MRC is in great attention of researchers due to its plentiful application scenar- ios, such as extractive and multi-choice MRC two major subcategories. Given a question with cor- responding passage, extractive MRC asks for pre- cise answer span extraction in passage (Joshi et al., ∗ Corresponding author. This paper was partially sup- ported by Key Projects of National Natural Science Founda- tion of China under Grants U1836222 and 61733011. Multi-choice MRC Example ... In addition, Lynn’s pioneering efforts also provide public educational forums via Green Scenes – a series of three hour events, each focusing on specific topics teaching Hoosiers how to lead greener lifestyles. ... Q: What can we learn about Green Scenes? A. It is a scene set in a three-hour film. B. It is a series of events focusing on green life. (Golden) C. It is a film set in Central Indiana. D. It is a forum focusing on green lifestyle. Extractive MRC Example ... Early versions were in use by 1851, but the most success- ful indicator was developed for the high speed engine inven- tor and manufacturer Charles Porter by Charles Richard and exhibited at London Exhibition in 1862. ... Q: Where was the Charles Porter steam engine indicator shown? Golden Answer: London Exhibition Imprecise Answer …

Content truncated to 500 words