Towards Summarizing Healthcare Questions in Low-Resource Setting

Proceedings of the 29th International Conference on Computational Linguistics, pages 2892–2905 October 12–17, 2022. 2892 Towards Summarizing Healthcare Questions in Low-Resource Setting Shweta Yadav and Cornelia Caragea Department of Computer Science, University of Illinois at Chicago, Illinois {shwetay, cornelia}@uic.edu Abstract The current advancement in abstractive docu- ment summarization depends to a large extent on a considerable amount of human-annotated datasets. However, the creation of large-scale datasets is often not feasible in closed domains, such as medical and healthcare domains, where human annotation requires domain expertise. This paper presents a novel data selection strat- egy to generate diverse and semantic questions in a low-resource setting with the aim to sum- marize healthcare questions. Our method ex- ploits the concept of guided semantic-overlap and diversity-based objective functions to opti- mally select the informative and diverse set of synthetic samples for data augmentation. Our extensive experiments on benchmark health- care question summarization datasets demon- strate the effectiveness of our proposed data selection strategy by achieving new state-of- the-art results. Our human evaluation shows that our method generates diverse, fluent, and informative summarized questions. 1 Introduction Online health information search is becoming con- ventional for more and more consumers every day. A recent survey showed that on average eight mil- lion people in the United States seek health-related information on the Internet every day1. One chal- lenge towards assisting consumers in their health- care information search is automatic question un- derstanding. Generally consumers’ questions are overly descriptive and include several peripheral information (as shown in Figure-1), which are not necessary to answer questions. Therefore, in this study we tackle the task of consumer health ques- tion understanding by summarizing the question. Automatic text summarization is a non-trivial task in Natural Language Processing (NLP) that aims to generate human-readable, concise text con- 1https://pewrsr.ch/3l6m3mv Figure 1: The highlighted text shows important key aspects of the question which need to be considered while generating the summary. taining salient information of the original docu- ment. The recent development in large-scale neural language models (Devlin et al., 2019; Raffel et al., 2020) have led to significant performance on sev- eral abstractive summarization task. However, their accuracy is partially due to the availability of large- scale human-annotated training data. Moreover, some domains such as biomedical and medical re- quire domain experts to create high-quality training datasets, which is tedious to create at a large-scale level. A potential solution that has shown effectiveness in other generation and translation tasks is to aug- ment the large-scale synthetically generated sam- ples with a human-annotated training set. However, a limited study focused on data selection strategy in summarization, particularly for abstractive summa- rization. The majority of the traditional data selec- tion methods are based on word replacement that mainly generates a synthetic sentence by changing one or multiple words with their synonyms (Zhang et al., 2015) or with a language model predicted words (Kobayashi, 2018). However, these methods make minor changes to the original sentence and therefore fall short of generating a diversified sen- …

Content truncated to 500 words