Clues Before Answers: Generation-Enhanced Multiple-Choice QA

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3272 - 3287 July 10-15, 2022 ©2022 Association for Computational Linguistics Clues Before Answers: Generation-Enhanced Multiple-Choice QA Zixian Huang and Ao Wu and Jiaying Zhou State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China {zixianhuang,awu,jyzhou}@smail.nju.edu.cn Yu Gu The Ohio State University, Columbus, USA gu.826@osu.edu Yue Zhao and Gong Cheng State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China yuezhao@smail.nju.edu.cn, gcheng@nju.edu.cn Abstract A trending paradigm for multiple-choice ques- tion answering (MCQA) is using a text-to-text framework. By unifying data in different tasks into a single text-to-text format, it trains a gen- erative encoder-decoder model which is both powerful and universal. However, a side ef- fect of twisting a generation target to fit the classification nature of MCQA is the under- utilization of the decoder and the knowledge that can be decoded. To exploit the genera- tion capability and underlying knowledge of a pre-trained encoder-decoder model, in this pa- per, we propose a generation-enhanced MCQA model named GenMC. It generates a clue from the question and then leverages the clue to en- hance a reader for MCQA. It outperforms text- to-text models on multiple MCQA datasets. 1 Introduction Multiple-choice question answering (MCQA) aims at selecting the correct answer from a set of options given a question. This long-standing challenge in natural language processing (NLP) requires ma- chines to have a wealth of knowledge, such as commonsense knowledge (Talmor et al., 2019; Mi- haylov et al., 2018) and scientific knowledge (Clark et al., 2018; Khot et al., 2020; Huang et al., 2019; Li et al., 2021), and have reasoning skills such as multi-hop reasoning (Khot et al., 2019) and logical reasoning (Yu et al., 2020; Liu et al., 2020b; Li et al., 2022). MCQA has made great progress with the devel- opment of pre-trained language models (PLMs). Basically there are two types of PLMs that are suit- able for different tasks. BERT (Devlin et al., 2019) and its variants such as RoBERTa (Liu et al., 2019) and ALBERT (Lan et al., 2020) are encoder-only models, being more suitable for natural language understanding (NLU) tasks including MCQA and other classification and regression tasks. T5 (Raffel et al., 2020) and BART (Lewis et al., 2020) are encoder-decoder models, being more suitable for natural language generation (NLG) tasks. How- ever, encoder-decoder models can also be applied to MCQA (Khashabi et al., 2020; Zhou et al., 2021). This is enabled by the text-to-text frame- work, which transforms data in different tasks into a unified text-to-text format so that knowledge span- ning many and various tasks can be learned, aggre- gated, and used by a single model. Research Question To fit MCQA, existing im- plementations of the text-to-text framework take all the options as input and are trained to gener- ate one of the options, i.e., to copy some tokens from the input. However, this is inconsistent with how encoder-decoder models are pre-trained so …

Content truncated to 500 words