A Corpus of Very Short Scientific Summaries

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 153–164 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 153 A Corpus of Very Short Scientific Summaries Yifan Chen University of Cambridge, UK yc462@cam.ac.uk Tamara Polajnar Royal Society of Chemistry and University of Cambridge, UK polajnart@rsc.org Colin Batchelor Royal Society of Chemistry, UK batchelorc@rsc.org Simone Teufel University of Cambridge, UK sht25@cam.ac.uk Abstract We present a new summarisation task, taking scientific articles and producing journal table- of-contents entries in the chemistry domain. These are one- or two-sentence author-written summaries that present the key findings of a pa- per. This is a first look at this summarisation task with an open access publication corpus consisting of titles and abstracts, as input texts, and short author-written advertising blurbs, as the ground truth. We introduce the dataset and evaluate it with state-of-the-art summarisation methods. 1 Introduction Table-of-contents (TOC) entries are short sum- maries written by authors that are placed in the table of contents of journals, often with an eye- catching accompanying image, to advertise their paper to readers. They are meant to be a clear and concise summary of a paper’s main contribu- tion, but different from the title and abstract, which have a different communicative function. We take the titles and abstracts from chemistry papers pub- lished by the Royal Society of Chemistry as input in this initial study, as they are freely available and more numerous, but will also release full text for the smaller subset of open access publications. As such, this particular corpus is different from other scientific corpora, which normally take the abstract as the summary. We include an analysis of the corpus proper- ties, and experimental results with strong baselines and three different state-of-the-art deep learning models: an attention-based RNN, a reinforcement learning approach, and a BERT-based transformer method. We also perform a human evaluation study to compare the models and to validate the useful- ness of the summarisation task. The main aim of releasing this particular corpus is to see how useful deep learning models are in producing a latent semantic representation of chem- ical texts. Within this paper we constrain ourselves to the question: Can we decode this representation into a useful summary and perhaps aid in the jour- nal editing process? But the real goal would be to transfer this representation into useful editing tasks like plagiarism detection or semantic search and discovery. In addition, chemistry has a complex domain lex- icon involving a potentially infinite set of molecules that can be described both by formulae and (of- ten multi-word) terms, as well as a set of other techniques. While everything from tokenisation onward in an NLP pipeline would benefit from customisation, there is a lack of domain-specific resources. In this paper we start addressing this problem by introducing a summarisation corpus, but we hope this is the first of many resources that will aid researchers in this field. Ultimately, un- derlying all these tasks is an ability to produce representations …

Content truncated to 500 words