Automatic related work section generation: experiments in scientific document abstracting

• Published in: Scientometrics Vol. 125; no. 3; pp. 3159 - 3185
• Main Authors: AbuRa’ed, Ahmed, Saggion, Horacio, Shvets, Alexander, Bravo, Àlex
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01-12-2020
• Subjects: Computer Science; Information Storage and Retrieval; Library Science; Information extraction from scientific literature; Scientific summarization; Sequence learning; Document abstracting; I.2.7; 97R40; 68T50
• ISSN: 0138-9130; 1588-2861
• DOI: 10.1007/s11192-020-03630-2

Related work sections or literature reviews are an essential part of every scientific article being crucial for paper reviewing and assessment. However, writing a good related work section is an activity which requires considerable expertise to identify, condense/summarize, and combine relevant information from different sources. In this work we compare different automatic methods to produce “descriptive” related work sections given as input the set of papers which need to be described. The main contribution of our work is a neural sequence learning process which produces citation sentences to be included in a related work section of an article. We train the neural architecture using an available scientific data set of citation sentences and we test over a data set of related work sections; we also compare the performance to a set of baseline extractive summarizers, an abstractive summarizer and a state of the art CNNs approach. Our results indicate that our approach outperforms the simple as well as the informed baselines.