NUT-RC: Noisy User-generated Text-oriented Reading Comprehension

Proceedings of the 28th International Conference on Computational Linguistics, pages 2687–2698 Barcelona, Spain (Online), December 8-13, 2020 2687 NUT-RC: Noisy User-generated Text-oriented Reading Comprehension Rongtao Huang1, Bowei Zou2∗, Yu Hong1, Wei Zhang3, Ai Ti Aw2, Guodong Zhou1 1School of Computer Scienceand Technology, Soochow University, China 2Aural & Language Intelligence Department, Institute for Infocomm Research, Singapore 3Alibaba Group, China rthuang.suda@gmail.com, {zou bowei,aaiti}@i2r.a-star.edu.sg, lantu.zw@alibaba-inc.com, {yhong,gdzhou}@suda.edu.cn Abstract Reading comprehension (RC) on social media such as Twitter is a critical and challenging task due to its noisy, informal, but informative nature. Most existing RC models are developed on formal datasets such as news articles and Wikipedia documents, which severely limit their per- formances when directly applied to the noisy and informal texts in social media. Moreover, these models only focus on a certain type of RC, extractive or generative, but ignore the integration of them. To well address these challenges, we come up with a noisy user-generated text-oriented RC model. In particular, we first introduce a set of text normalizers to transform the noisy and informal texts to the formal ones. Then, we integrate the extractive and the generative RC model by a multi-task learning mechanism and an answer selection module. Experimental results on TweetQA demonstrate that our NUT-RC model significantly outperforms the state-of-the-art so- cial media-oriented RC models. 1 Introduction Reading Comprehension (RC), which aims to answer questions by comprehending the contexts of given passages, is a frontier topic in natural language processing research. Recently, many RC models (Wang et al., 2018; Lin et al., 2018; Zhu et al., 2018; Joty et al., 2018; Weber et al., 2019) have been proposed and have achieved considerable successes. According to answer prediction methods, the RC models can be roughly divided into two major categories: extractive and generative. For an extractive RC model, the predicted answer is limited to be a consecutive span in the passage. For a generative RC model, it allows the answer to be free-text which can include novel words and phrases not appeared in passages. However, most of the existing RC models are developed for formal text, such as news articles (Her- mann et al., 2015; Trischler et al., 2017) and Wikipedia (Rajpurkar et al., 2016; Joshi et al., 2017), which severely limit their performances on Noisy User-generated Texts (NUT). Meanwhile, an increas- ing number of people are accustomed to getting real-time information via social media like Twitter. Recently, Xiong et al. (2019) propose a large-scale dataset for question answering over social media texts, TweetQA, which is constructed in a crowd-sourcing way, and recommended annotators to write answers in their own language. Table 1 shows an example from the TweetQA dataset. In order to answer the question “what is kerry thankful for?”, an RC model first needs to be positioned to “Thank u 4 opening this convo”, where “u, 4, convoy” is the abbreviation of “you, for, conversation”, respectively. Furthermore, “u” refers to “@InStyle”. Then the RC model can predict the answer “instyle opening the conversation”. Such an example indicates that question …

Content truncated to 500 words