A Robust Abstractive System for Cross-Lingual Summarization

Proceedings of NAACL-HLT 2019, pages 2025–2031 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2025 A Robust Abstractive System for Cross-Lingual Summarization Jessica Ouyang and Boya Song and Kathleen McKeown Deptartment of Computer Science Columbia University New York, NY 10027 {ouyangj, kathy}@cs.columbia.edu, bs3065@columbia.edu Abstract We present a robust neural abstractive summa- rization system for cross-lingual summariza- tion. We construct summarization corpora for documents automatically translated from three low-resource languages, Somali, Swahili, and Tagalog, using machine translation and the New York Times summarization corpus. We train three language-specific abstractive sum- marizers and evaluate on documents originally written in the source languages, as well as on a fourth, unseen language: Arabic. Our sys- tems achieve significantly higher fluency than a standard copy-attention summarizer on auto- matically translated input documents, as well as comparable content selection. 1 Introduction Cross-lingual summarization is a little-explored task combining the difficulties of automatic sum- marization with those of machine translation. The goal is to summarize in one language a docu- ment available only in another language. Wan et al. (2010) describe two approaches: summarize then translate, and translate then summarize. They argue that summarize-then-translate is preferable to avoid both the computational expense of trans- lating more sentences and sentence extraction er- rors caused by incorrect translations. However, summarize-then-translate can only be used when the source language is high-resource (Wan et al. used English as the source, for exam- ple); if the source language is one of the thousands of low-resource languages in the world, there are no summarization corpora available. Language- independent techniques, such as TextRank (Mihal- cea), might be used, but there may be serious dif- ficulties in their application, such as morphologi- cally rich languages that render token-based simi- larity measures useless. In such a case, translate- then-summarize is the only possible approach. We address this scenario through the develop- ment of a neural abstractive summarization sys- tem that fluently summarizes potentially disfluent, automatically-translated documents by generating short, simple phrases to replace awkward input phrases resulting from difficult to translate source documents. Our novel combination of existing building block systems results in a summarization solution that can be easily applied to new low- resource languages. We use machine translation on the New York Times annotated corpus of docu- ment/summary pairs to create summarization cor- pora for documents automatically translated from three low-resource languages, Somali, Swahili, and Tagalog. We use these corpora to train cross- lingual summarizers for these source languages, with English as the target. We also evaluate our systems on a fourth source language, Arabic. Our experiments show that our abstractive summariz- ers produce more fluent English summaries from automatically-translated documents, and that this improvement generalizes across source languages. Our main contributions are as follows: • We create summarization corpora for automat- ically translated Somali, Swahili, and Taga- log documents: noisy English input documents paired with clean English reference summaries. • We present a method for producing cross- lingual summarization systems for low resource languages where no summarization corpora …

Content truncated to 500 words