Cross-Lingual Open-Domain Question Answering with Answer Sentence Generation

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 337–353 November 20–23, 2022. ©2022 Association for Computational Linguistics 337 Cross-Lingual Open-Domain Question Answering with Answer Sentence Generation Benjamin Muller1∗, Luca Soldaini2†, Rik Koncel-Kedziorski3, Eric Lind3, Alessandro Moschitti3 1Inria, Paris, France 2Allen Istitute for AI 3Amazon Alexa AI benjamin.muller@inria.fr, lucas@allenai.org, {rikdz,ericlind,amosch}@amazon.com Abstract Open-Domain Generative Question Answering has achieved impressive performance in En- glish by combining document-level retrieval with answer generation. These approaches, which we refer to as GENQA, can generate complete sentences, effectively answering both factoid and non-factoid questions. In this pa- per, we extend GENQA to the multilingual and cross-lingual settings. For this purpose, we first introduce GEN-TYDIQA, an exten- sion of the TyDiQA dataset with well-formed and complete answers for Arabic, Bengali, En- glish, Japanese, and Russian. Based on GEN- TYDIQA, we design a cross-lingual genera- tive model that produces full-sentence answers by exploiting passages written in multiple lan- guages, including languages different from the question. Our cross-lingual generative system outperforms answer sentence selection base- lines for all 5 languages and monolingual gen- erative pipelines for three out of five languages studied. 1 Introduction Improving coverage of the world’s languages is essential for retrieval-based Question Answering (QA) systems to provide a better experience for non-English speaking users. One promising direc- tion for improving coverage is multilingual, multi- source, open-domain QA. Multilingual QA sys- tems include diverse viewpoints by leveraging an- swers from multiple linguistic communities. Fur- ther, they can improve accuracy, as all facets nec- essary to answer a question are often unequally distributed across languages on the Internet (Valen- tim et al., 2021). With the advance of large-scale language mod- els, multilingual modeling has made impressive progress at performing complex NLP tasks without requiring explicitly translated data. Building on ∗ Work conducted during internship at Amazon Alexa. † Work conducted while employed at Amazon Alexa. pre-trained language models (Devlin et al., 2019; Conneau et al., 2020; Xue et al., 2021; Liu et al., 2020), it is now possible to train models that ac- curately process textual data in multiple languages (Kondratyuk and Straka, 2019) and perform cross- lingual transfer (Pires et al., 2019) using annotated data in one language to process another language. At the same time, answer generation-based ap- proaches have been shown to be effective for many English QA tasks, including Machine Read- ing (MR) (Izacard and Grave, 2021; Lewis et al., 2020c), question-based summarization (Iida et al., 2019; Goodwin et al., 2020; Deng et al., 2020), and, most relevant to this work, answer generation for retrieval-based QA (Hsu et al., 2021) — that we refer to as GENQA. Compared to generative MR models, GENQA approaches are trained to produce complete and expressive sentences that are easier to understand than extracted snippets (Choi et al., 2021). Most importantly, they are trained to generate entire sen- tences, allowing them to answer both factoid or non-factoid questions, …

Content truncated to 500 words