Attend, Translate and Summarize: An Efficient Method for Neural Cross-Lingual Summarization

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1309–1321 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1309 Attend, Translate and Summarize: An Efficient Method for Neural Cross-Lingual Summarization Junnan Zhu1,2, Yu Zhou1,2,3, Jiajun Zhang1,2, Chengqing Zong1,2 1 National Laboratory of Pattern Recognition, Institute of Automation, CAS 2 School of Artificial Intelligence, University of Chinese Academy of Sciences 3 Beijing Fanyu Technology Co., Ltd {junnan.zhu, yzhou, jjzhang, cqzong}@nlpr.ia.ac.cn Abstract Cross-lingual summarization aims at summa- rizing a document in one language (e.g., Chi- nese) into another language (e.g., English). In this paper, we propose a novel method inspired by the translation pattern in the process of ob- taining a cross-lingual summary. We first at- tend to some words in the source text, then translate them into the target language, and summarize to get the final summary. Specif- ically, we first employ the encoder-decoder attention distribution to attend to the source words. Second, we present three strategies to acquire the translation probability, which helps obtain the translation candidates for each source word. Finally, each summary word is generated either from the neural distribution or from the translation candidates of source words. Experimental results on Chinese-to- English and English-to-Chinese summariza- tion tasks have shown that our proposed method can significantly outperform the base- lines, achieving comparable performance with the state-of-the-art. 1 Introduction Cross-lingual summarization is to produce a sum- mary in a target language (e.g., English) from a doc- ument in a different source language (e.g., Chinese). Cross-lingual summarization can help people effi- ciently understand the gist of an article written in an unfamiliar foreign language. Traditional cross-lingual summarization meth- ods are pipeline-based. These methods either adopt summarize-then-translate (Orasan and Chiorean, 2008; Wan et al., 2010) or employ translate-then- summarize (Leuski et al., 2003; Ouyang et al., 2019). The pipeline-based approach is intuitive and straightforward, but it suffers from error propaga- tion. Due to the difficulty of acquiring cross-lingual summarization dataset, some previous researches focus on zero-shot methods (Ayana et al., 2018; �����������������������10��������� ��������������������������������� ��������������������������������� Foshan young couple was detained for charging 10 yuan for buying train tickets online for migrant workers A young couple in Foshan who help migrant workers book train tickets online have been detained after receiving a 10-yuan handling fee for each ticket. Migrant workers called injustice, and they did not overcharge, much better than scalpers. Lawyers said that under the law, disguised mark-up scalping and other acts constitute scalping tickets serious, will be punished. � �—Foshan ���—young couple ���—train tickets 10 �—10 yuan � �—detained ���—migrant workers Input (Chinese) Output (English) Input (English) Translation Table Figure 1: An example of the translation pattern in a sample extracted from Zh2EnSum (Zhu et al., 2019) which is a Chinese-to-English cross-lingual summa- rization dataset. It shows that some words in the sum- mary are translated from the source words (in the same color). The translation table also gives the correspond- ing relation to these words. Best viewed in color. Duan et al., 2019), i.e., …

Content truncated to 500 words