Emerald 110k: A Multidisciplinary Dataset for Abstract Sentence Classification

Abstract Background: Datasets available for abstract sentence classification modelling are predominately comprised of abstracts sourced from biomedical research. Aims: To contribute a large non-biomedical multidisciplinary dataset for abstract sentence classification model research. Method: Bulk extract and transformation of Emerald Group Publishing structured abstracts indexed on Scopus. Results: We present the largest multidisciplinary dataset for abstract sentence classification modelling, consisting of 1,050,397 sentences from 103,457 abstracts. 1 Introduction Abstracts enable researchers to efficiently determine the relevance of literature to their research (Rowley, 1982, Collision, 1971, Cleveland and Cleveland, 2013). The desire to optimise this efficiency has resulted in the adoption of structured abstracts, which feature explicit headings reflecting key characteristics of a study. Examples of these headings include: aim, method, results and contributions. The alternative to structured abstracts are those where sentences addressing such characteristics are not specified. Compared to unstructured alternatives, structured abstracts are perceived to offer greater value for researchers (Sharma and Harrison, 2006, Taddio et al., 1994 and Guimarães, 2006); permit advanced access to research findings (Mosteller et al., 2004), contain more relevant information (Budgen et al., 2008) and are easier to read (Kitchenham et al., 2008 and Budgen et al., 2008). Structured abstracts also increase the likelihood that relevant research is discovered (Eldredge, 2006, Mulrow, 1987, Haynes et al., 1990, Hartley, 1997, Bayley et al., 2002 and Bayley and Eldredge, 2003). Natural language processing (NLP) has been used to automate the structuring of unstructured abstracts (Gonçalves et al., 2018; Jin and Szolovits, 2018, Dernoncourt et al., 2016); which is achieved through the development of Abstract Sentence Classification Models (ASCM), capable of classifying sentences sourced from unstructured abstracts into structured abstract headings. This paper presents a novel dataset to advance ASCM research. The dataset introduced is unlike those already leveraged in ASCM development, primarily as it is comprised of abstracts originating from disciplines not yet explored in current research. The adoption of our dataset in future model development will enable the benchmarking of ASCM capability in new disciplines. 2 Related Work There are numerous datasets available to researchers seeking to develop ASCM. These are outlined in table 1, an extension of the table presented by Dernoncourt and Lee (2017, p. 3). We extend their table by identifying the abstract’s disciplinary domain. The size represents the number of abstracts reflected in the dataset. The ‘manual’ flag identifies if sentences were manually classified into structured abstract headings by the authors (Y) or were pre-structured in the original abstract (N). The number of datasets available for ASCM does not directly correspond to the number of ASCM studies, as researchers re-use datasets to benchmark performance and to test novel algorithms. Further, studies may develop a dataset for model development without contributing the dataset as an artefact. Dernoncourt and Lee (2017) also presented a dataset in a standalone paper, Emerald 110k: A Multidisciplinary Dataset for Abstract Sentence Classification Connor Stead 1 Stephen Smith 1 Peter Busch 1 Savanid Vatanasakdakul 2 1 Department of Computing, Macquarie University, Australia 2 Carnegie Mellon University, Qatar connor.stead@hdr.mq.edu.au …

Content truncated to 500 words