DaNewsroom: A Large-scale Danish Summarisation Dataset

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6731–6739 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6731 DaNewsroom: A Large-scale Danish Summarisation Dataset Daniel Varab and Natalie Schluter IT University of Copenhagen Copenhagen, Denmark {djam, natschluter}@itu.dk Abstract Dataset development for automatic summarisation systems is notoriously English-oriented. In this paper we present the first large-scale non-English language dataset specifically curated for automatic summarisation. The document-summary pairs are news articles and manually written summaries in the Danish language. There has previously been no work done to establish a Danish summarisation dataset, nor any published work on the automatic summarisation of Danish. We provide therefore the first automatic summarisation dataset for the Danish language (large-scale or otherwise). To support the comparison of future automatic summarisation systems for Danish, we include system performance on this dataset of strong well-established unsupervised baseline systems, together with an oracle extractive summariser, which is the first account of automatic summarisation system performance for Danish. Finally, we make all code for automatically acquiring the data freely available and make explicit how this technology can easily be adapted in order to acquire automatic summarisation datasets for further languages. Keywords: automatic text summarisation, data collection, danish corpus 1 Introduction Dataset development for automatic summarisation systems is notoriously English-oriented. This is surprising. On the system user-side, a more feasible access (for example, summaries) to the increasing amounts of digital information informing daily life is of inherent interest to potential users across the globe. At the same time, automatic summari- sation provides a challenging NLP test-bed to investigate the limits of deep learning for NLP, and for downstream evaluation of basic core NLP tasks like discourse analysis, co-reference resolution, and other types of parsing (Lee et al., 2018; Li et al., 2016). Yet, only very limited datasets ex- ist in languages other than English (Nguyen and Daumé III, 2019; Schluter and Martínez Alónso, 2016). By automatic summarisation dataset we denote a collec- tion of entire documents each paired up with at least one manually written summary; the summaries of such a dataset are intended as a summaries for those documents and not as headlines, or a list of facts or highlights. In fact, un- til recently central larger-scale automatic summarisation datasets have not included been composed of any sum- maries. Namely, Rush et al. (Rush et al., 2015) were the first to recast the English Gigaword dataset (Parker et al., 2011; Napoles et al., 2012) as a headline-type large-scale summarisation generation dataset. And the CNN/Daily Mail dataset (Moritz Hermann et al., 2015), a question answering dataset, was first recast by (Cheng and Lapata, 2016; Nallapati et al., 2016) as an automatic summarisation dataset. These two datasets have been central to more recent automatic summarisation system development. Headline and highlights datasets are are not ideal for the de- velopment of summarisation systems, but because of their scale and in the absence of alternatives, they provided a much needed crucial prerequisite for neural system …

Content truncated to 500 words