Enhancing Answer Boundary Detection for Multilingual Machine Reading Comprehension

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 925–934 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 925 Enhancing Answer Boundary Detection for Multilingual Machine Reading Comprehension Fei Yuan1∗ Linjun Shou2† Xuanyu Bai2 Ming Gong2 Yaobo Liang3 Nan Duan3 Yan Fu1 Daxin Jiang2 1University of Electronic Science and Technology of China 2STCA NLP Group, Microsoft, Beijing, China 3Microsoft Research Asia, Beijing, China feiyuan@std.uestc.edu.cn {lisho,xub,migon,yalia,nanduan,djiang}@microsoft.com fuyan@uestc.edu.cn Abstract Multilingual pre-trained models could lever- age the training data from a rich source lan- guage (such as English) to improve the perfor- mance on low resource languages. However, the transfer effectiveness on the multilingual Machine Reading Comprehension (MRC) task is substantially poorer than that for sentence classification tasks, mainly due to the require- ment of MRC to detect the word level an- swer boundary. In this paper, we propose two auxiliary tasks to introduce additional phrase boundary supervision in the fine-tuning stage: (1) a mixed MRC task, which translates the question or passage to other languages and builds cross-lingual question-passage pairs; and (2) a language-agnostic knowledge mask- ing task by leveraging knowledge phrases mined from the Web. Extensive experiments on two cross-lingual MRC datasets show the effectiveness of our proposed approach. 1 Introduction Machine Reading Comprehension (MRC) plays a critical role in the assessment of how well a ma- chine could understand natural language. Among various types of MRC tasks, the span extractive reading comprehension task (like SQuAD (Ra- jpurkar et al., 2016)) has been become very pop- ular. Promising achievements have been made with neural network based approaches (Seo et al., 2017; Wang et al., 2017; Xiong et al., 2018; Yu et al., 2018; Hu et al., 2017), especially those built on pre-trained language models such as BERT (Devlin et al., 2018), due to the availability of large-scale annotated corpora (Hermann et al., 2015; Rajpurkar et al., 2016; Joshi et al., 2017). However, these large-scale annotated corpora are ∗Work is done during internship at STCA NLP Group, Microsoft. †Correspondence author. Language MRC NLI EM Gap to en ACC Gap to en en 62.4 – 85.0 – es 49.8 -12.6 78.9 -6.1 de 47.6 -14.8 77.8 -7.2 ar 36.3 -26.1 73.1 -11.9 hi 27.3 -35.1 69.6 -15.4 vi 41.8 -20.6 76.1 -8.9 zh 39.6 -22.8 76.5 -8.5 Table 1: The gap between target languages and English on Machine Reading Comprehension (MRC) (Lewis et al., 2019) is significantly larger than sentence level classification task like Natural Language Inference (NLI) (Conneau et al., 2018). In this experiment, we fine-tune XLM (Conneau and Lample, 2019) on En- glish and directly test on other languages. mostly exclusive to English, while research about MRC on languages other than English (i.e. multi- lingual MRC) has been limited due to the absence of sufficient training data. To alleviate the scarcity of training data for mul- tilingual MRC, the translation based data augmen- tation approaches were firstly proposed. For ex- ample, (question q, passage p, answer a) in En- glish SQuAD can be translated into …

Content truncated to 500 words