A Hybrid Approach to Generation of Missing Abstracts in Biomedical Literature

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1093–1100, Osaka, Japan, December 11-17 2016. A Hybrid Approach to Generation of Missing Abstracts in Biomedical Literature Suchet K Chachra suchet.chachra@gmail.com Asma Ben Abacha asma.benabacha@nih.gov Sonya Shooshan sonya@nlm.nih.gov Laritza Rodriguez laritza.rodriguez@nih.gov Dina Demner-Fushman ddemner@mail.nih.gov U.S. National Library of Medicine, Bethesda, MD, USA Abstract Readers usually rely on abstracts to identify relevant medical information from scientific arti- cles. Abstracts are also essential to advanced information retrieval methods. More than 50 thou- sand scientific publications in PubMed Central lack author-generated abstracts, and the relevancy judgements for these papers have to be based on their titles alone. In this paper, we propose a hybrid summarization technique that aims to select the most pertinent sentences from articles to generate an extractive summary in lieu of a missing abstract. We combine i) health outcome detection, ii) keyphrase extraction, and iii) textual entailment recognition between sentences. We evaluate our hybrid approach and analyze the improvements of multi-factor summarization over techniques that rely on a single method, using a collection of 295 manually generated reference summaries. The obtained results show that the hybrid approach outperforms the baseline tech- niques with an improvement of 13% in recall and 4% in F1 score. 1 Introduction PubMed Central1 (PMC) is a repository of biomedical and life sciences journals supported by the U.S. National Library of Medicine (NLM). PMC provides access to the abstracts as well as the full-text content of biomedical articles. The open-access subset of PubMed Central contains over one million biomedical articles as of Fall 2015, and is widely used as a public resource to discover, read and build upon its vast portfolio of biomedical knowledge. Given the abundance and variety of information avail- able within PMC, many user queries return a multitude of results, which makes it more difficult to iden- tify relevant data. The amount of potentially relevant results often increases further due to information retrieval techniques such as query expansion using synonyms. Article abstracts are usually considered to be entry points into the full-text. Abstracts often contain key health-outcome data and clinical findings that help identifying relevant data when narrowing down the number of returned results to a select few. The abstracts, however, are missing in 50,000 articles available within the open access subset of PMC. Therefore, for this large set of articles, the only way for the users to judge the relevancy of an article is either through the article title, which is not always reliable, or through the full-text of the paper, which can be time-consuming. In this paper, we describe a novel hybrid approach that builds upon textual entailment, keyphrase extraction and health outcome detection to generate surrogate abstracts for biomedical articles where none are available. Using a set of 295 documents and manually generated extractive summaries that we make publicly available with this publication, we also analyze how this approach compares to baseline methods relying on a single technique. 2 Related Work Single and multi-document text summarization …

Content truncated to 500 words