LipKey: A Large-Scale News Dataset for Absent Keyphrases Generation and Abstractive Summarization

Proceedings of the 29th International Conference on Computational Linguistics, pages 3427–3437 October 12–17, 2022. 3427 LipKey: A Large-Scale News Dataset for Absent Keyphrases Generation and Abstractive Summarization Fajri Koto♠ Timothy Baldwin♠♥ Jey Han Lau♠ ♠ School of Computing and Information Systems, The University of Melbourne ♥ Department of Natural Language Processing, MBZUAI fajri.koto91@gmail.com tb@ldwin.net jeyhan.lau@gmail.com Abstract Summaries, keyphrases, and titles are differ- ent ways of concisely capturing the content of a document. While most previous work has released the datasets of keyphrases and summarization separately, in this work, we introduce LipKey, the largest news corpus with human-written abstractive summaries, ab- sent keyphrases, and titles. We jointly use the three elements via multi-task training and training as joint structured inputs, in the con- text of document summarization. We find that including absent keyphrases and titles as ad- ditional context to the source document im- proves transformer-based summarization mod- els.1 1 Introduction Key content of an article can be presented in dif- ferent ways, including summaries, keyphrases, and titles. While most previous research has addressed each element individually (e.g. news summariza- tion (Zhang et al., 2020a; Lewis et al., 2020; Koto et al., 2020a) and keyphrase generation in the sci- entific domain (Meng et al., 2017, 2021)), in this work we release a novel news dataset that consists of highly absent keyphrases (i.e. keyphrases which abstract over the content of the document), abstrac- tive summaries, and titles to investigate the three elements in the context of single-document abstrac- tive summarization. Previous work has mainly utilized present keyphrases (i.e. keyphrases that are directly drawn from the source text) through unsupervised and supervised methods for summarization. For in- stance, traditional summarization models (Zhang et al., 2004; D’Avanzo and Magnini, 2005; Wan et al., 2007; Riedhammer et al., 2010; Qazvinian et al., 2010) and modern neural models (Müngen and Kaya, 2018; Nallapati et al., 2016; Liu et al., 1Data and code used is available at: https://github.com/ fajri91/LipKey 2021) have been combined with the top-k frequent words, TF-IDF, and TextRank (Mihalcea and Tarau, 2004) to obtain keyphrases. Elsewhere, Gehrmann et al. (2018); Li et al. (2020) used words contained in both the summary and article as keyphrases to improve summarization. This paper aims to study how absent keyphrases (i.e. keyphrases that do not match any words in the source text) can be incorporated into summa- rization systems. Compared to present keyphrases used in previous work, absent keyphrases poten- tially better complement abstractive summariza- tion methods. Previous work has been hindered by the unavailability of a large annotated dataset with gold-standard summaries and keyphrases, thus opting for present keyphrase extraction (Qazvinian et al., 2010; Liu et al., 2021). We additionally study the utility of titles in sum- marization. The underlying hypothesis is that titles and keyphrases are concise, complementary repre- sentations of an article, and provide relevant clues for summarization. While previous summarization datasets such as CNNDM (Hermann et al., 2015), NYT (Sandhaus, 2008), and XSUM (Narayan et al., 2018) do not include keyphrases and titles, we …

Content truncated to 500 words