Constrained Regeneration for Cross-Lingual Query-Focused Extractive Summarization

Proceedings of the 29th International Conference on Computational Linguistics, pages 2668–2680 October 12–17, 2022. 2668 Constrained Regeneration for Cross-Lingual Query-Focused Extractive Summarization Elsbeth Turcan1, David Wan2,∗, Faisal Ladhak1, Petra Galuscakova3,†, Sukanta Sen4,‡ Svetlana Tchistiakova4,‡, Weijia Xu5, Marine Carpuat5, Kenneth Heafield4, Douglas Oard5 and Kathleen McKeown1 1Department of Computer Science, Columbia University; 2Department of Computer Science, University of North Carolina at Chapel Hill; 3Université Grenoble Alpes; 4School of Informatics, The University of Edinburgh; 5Department of Computer Science & UMIACS, The University of Maryland {eturcan, faisal}@cs.columbia.edu, davidwan@cs.unc.edu Abstract Query-focused summaries of foreign-language, retrieved documents can help a user under- stand whether a document is actually relevant to the query term. A standard approach to this problem is to first translate the source docu- ments and then perform extractive summariza- tion to find relevant snippets. However, in a cross-lingual setting, the query term does not necessarily appear in the translations of relevant documents. In this work, we show that constrained machine translation and con- strained post-editing can improve human rel- evance judgments by including a query term in a summary when its translation appears in the source document. We also present several strategies for selecting only certain documents for regeneration which yield further improve- ments. 1 Introduction Query-focused summarization creates an overview of a document which reflects how that document may be relevant to a provided query; such a task is useful for any search engine, such as for news articles or academic papers, where a user may want to search documents by a given query. In this paper, we further narrow the use case to one in which the user seeks a document containing a single specific query term (which may be a multi-word expres- sion). For example, if the query is “dossier”, the user is interested in finding information about a specific type of collection of files, as might exist in an intelligence investigation. A summary in the user’s language can help them decide if a foreign language document is relevant. Our work focuses on query-focused extractive summarization in a cross-lingual setting, where the ∗Work performed while at Columbia University †Work performed while at the University of Maryland ‡Work performed while at the University of Edinburgh summaries are generated in a language (here, En- glish) different from the source language of the documents (here, Farsi, Kazakh, and Georgian). Because large summarization corpora do not ex- ist in these languages, we follow a translate-then- summarize approach (Wan et al., 2010) in which we first apply machine translation (MT) to translate documents into English, a language with abundant summarization corpora, and then summarize the translated document; however, this introduces ad- ditional concerns. Translating a document once, before a query term is known, can lead to word- ing choices that are sub-optimal for any particular query term (e.g., if the Kazakh for “dossier” were translated as “file”, and it may be unclear whether the specific meaning of dossier occurred in the source as opposed to other meanings of “file”). To address this, we present …

Content truncated to 500 words