Cross-language Sentence Selection via Data Augmentation and Rationale Training

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3881–3895 August 1–6, 2021. ©2021 Association for Computational Linguistics 3881 Cross-language Sentence Selection via Data Augmentation and Rationale Training Yanda Chen1, Chris Kedzie1, Suraj Nair2, Petra Galuˇsˇc´akov´a2, Rui Zhang3, Douglas W. Oard2, Kathleen McKeown1 1Columbia University, 2University of Maryland, 3Penn State University yc3384@columbia.edu, {kedzie, kathy}@cs.columbia.edu {srnair, petra, oard}@umd.edu, rmz5227@psu.edu Abstract This paper proposes an approach to cross- language sentence selection in a low-resource setting. It uses data augmentation and neg- ative sampling techniques on noisy parallel sentence data to directly learn a cross-lingual embedding-based query relevance model. Re- sults show that this approach performs as well as or better than multiple state-of-the- art machine translation + monolingual re- trieval systems trained on the same paral- lel data. Moreover, when a rationale train- ing secondary objective is applied to encour- age the model to match word alignment hints from a phrase-based statistical machine trans- lation model, consistent improvements are seen across three language pairs (English- Somali, English-Swahili and English-Tagalog) over a variety of state-of-the-art baselines. 1 Introduction Sentence-level query relevance prediction is impor- tant for downstream tasks such as query-focused summarization and open-domain question answer- ing; accurately pinpointing sentences containing information that is relevant to the query is critical to generating a responsive summary/answer (e.g., Baumel et al. (2016, 2018)). In this work, we fo- cus on sentence-level query relevance prediction in a cross-lingual setting, where the query and sen- tence collection are in different languages and the sentence collection is drawn from a low-resource language. Our approach enables English speakers (e.g., journalists) to find relevant information ex- pressed in local sources (e.g., local reaction to the pandemic and vaccines in Somalia). While we can use machine translation (MT) to translate either the query or each sentence into a common language, and then use a monolingual In- formation Retrieval (IR) system to find relevant sen- tences, work on Probabilistic Structured Queries (PSQ) (Darwish and Oard, 2003) has shown that the performance of such MT+IR pipelines is hin- dered by errors in MT. As is well known, complete translation of the sentence collection is not neces- sary. Inspired by previous work (Vuli´c and Moens, 2015), we go a step further and propose a simple cross-lingual embedding-based model that avoids translation entirely and directly predicts the rele- vance of a query-sentence pair (where the query and sentence are in different languages). For training, we treat a sentence as relevant to a query if there exists a translation equivalent of the query in the sentence. Our definition of relevance is most similar to the lexical-based relevance used in Gupta et al. (2007) and Baumel et al. (2018) but our query and sentence are from different languages. We frame the task as a problem of finding sen- tences that are relevant to an input query, and thus, we need relevance judgments for query-sentence pairs. Our focus, however, is on low-resource lan- guages …

Content truncated to 500 words