Cross-Lingual Question Answering over Knowledge Base as Reading Comprehension

Findings of the Association for Computational Linguistics: EACL 2023, pages 2439–2452 May 2-6, 2023 ©2023 Association for Computational Linguistics Cross-Lingual Question Answering over Knowledge Base as Reading Comprehension Chen Zhang1, Yuxuan Lai3, Yansong Feng1,2∗, Xingyu Shen1, Haowei Du1, Dongyan Zhao1,4,5 1 Wangxuan Institute of Computer Technology, Peking University, China 2 The MOE Key Laboratory of Computational Linguistics, Peking University, China 3 Department of Computer Science, The Open University of China 4 National Key Laboratory of General Artificial Intelligence 5 Beijing Institute for General Artificial Intelligence {zhangch,fengyansong,shenxy,zhaody}@pku.edu.cn duhaowei@stu.pku.edu.cn laiyx@ochn.edu.cn Abstract Although many large-scale knowledge bases (KBs) claim to contain multilingual informa- tion, their support for many non-English lan- guages is often incomplete. This incomplete- ness gives birth to the task of cross-lingual ques- tion answering over knowledge base (xKBQA), which aims to answer questions in languages different from that of the provided KB. One of the major challenges facing xKBQA is the high cost of data annotation, leading to limited resources available for further exploration. An- other challenge is mapping KB schemas and natural language expressions in the questions under cross-lingual settings. In this paper, we propose a novel approach for xKBQA in a read- ing comprehension paradigm. We convert KB subgraphs into passages to narrow the gap be- tween KB schemas and questions, which en- ables our model to benefit from recent advances in multilingual pre-trained language models (MPLMs) and cross-lingual machine reading comprehension (xMRC). Specifically, we use MPLMs, with considerable knowledge of cross- lingual mappings, for cross-lingual reading comprehension. Existing high-quality xMRC datasets can be further utilized to finetune our model, greatly alleviating the data scarcity is- sue in xKBQA. Extensive experiments on two xKBQA datasets in 12 languages show that our approach outperforms various baselines and achieves strong few-shot and zero-shot perfor- mance. Our dataset and code are released for further research1. 1 Introduction Large-scale knowledge bases (KBs) such as Free- base (Bollacker et al., 2008) and DBpedia (Auer et al., 2007) store huge amounts of structured ∗Corresponding author. 1https://github.com/luciusssss/xkbqa-as-mrc Marvel Universe appears in universes Knowledge Base (Subgraph): Question: ������������� ��������������������������� Qui joue Mary Jane dans Spiderman 2? �Who plays Mary Jane in Spiderman 2?� Answer Stan Lee character created by The Punisher Strikes Twice story appearances Mary Jane Watson Spider- man 2 Kirsten Dunst CVT portrayed in films film actor Figure 1: An example of answering questions in non- English languages over an English knowledge base. knowledge. These KBs support a variety of natu- ral language processing tasks, including question answering over knowledge base (KBQA), where models exploit the knowledge related to the ques- tions and precisely identify the answers by rea- soning through various KB relations. Although most large-scale KBs claim to contain multilin- gual information, they could not completely sup- port non-English languages as expected. For ex- ample, Freebase has no translation for the KB relations/attributes in any non-English languages. More than half of the entities in Freebase have no Chinese translations, despite the fact that Chinese is the most spoken non-English language in the …

Content truncated to 500 words