Impact of Citing Papers for Summarisation of Clinical Documents

Impact of Citing Papers for Summarisation of Clinical Documents Diego Moll´a Christopher Jones Macquarie University Sydney, Australia diego.molla-aliod@mq.edu.au christopher.jones4@students.mq.edu.au Abeed Sarker Arizona State University Tempe, AZ, USA abeed.sarker@asu.edu Abstract In this paper we show that information from citing papers can help perform ex- tractive summarisation of medical publica- tions, especially when the amount of text available for development is limited. We used the data of the TAC 2014 biomed- ical summarisation task. We report sev- eral methods to find the reference pa- per sentences that best match the citation text from the citing papers (“citances”). We observed that methods that incorporate lexical domain information from UMLS, and methods that use extended training data, perform best. We then used these ranked sentences to perform extractive summarisation and observed a dramatic improvement of ROUGE-L scores when compared with methods that do not use in- formation from citing papers. 1 Introduction Text-based summarisation is a well-established area of research that aims to automatically pro- duce condensed text representations of the origi- nal text. Text-based summarisation is useful in an increasing number of application domains where people cannot afford to spend time to read all the relevant information. This is certainly the case in the medical domain, and several approaches for the automated summarisation of medical text have been proposed, e.g. as surveyed by Afantenos et al. (2005). Information from citing texts has been used in decades-old studies (Garfield et al., 1964). More recently, Nakov et al. (2004) proposed the use of citations for the semantic interpretation of bio- science text. They used the text surrounding the citations, which they named “citances”, to sum- marise the original text. Further research focused on the extraction of the citances and surrounding text (Qazvinian and Radev, 2010) and on the use of these citances to gather information about the original text, which could be used as a surrogate of, or in addition to, a summary of the text (Mo- hammad et al., 2009; Abu-Jbara and Radev, 2011). The Biomedical Summarization Track of the 2014 Text Analysis Conference (TAC 2014 BiomedSumm Track)1 was designed as a set of shared tasks that focus on the use of the ci- tances to build summaries of biomedical docu- ments. The track organisers provided a small data set of 20 biomedical documents for training and fine-tuning. Each paper of the data set (henceforth “reference paper”) has 10 citing papers, and the data are annotated with the citances found in the citing papers. For each citance, four annotators appointed by the National Institute of Standards and Technology (NIST) identified various pieces of information related to the track tasks. Three tasks were defined: Task 1a Identify the text spans from the refer- ence paper that most accurately reflect the text from the citance. Task 1b Classify what facet of the paper a text span belongs to. There are 6 fixed facets: hypothesis, method, results, implication, dis- cussion, and data-set-used. Task 2 Generate a structured summary of the ref- erence paper and all of the …

Content truncated to 500 words