Paragraph-level Simplification of Medical Texts

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4972–4984 June 6–11, 2021. ©2021 Association for Computational Linguistics 4972 Paragraph-level Simplification of Medical Texts Ashwin Devaraj The University of Texas at Austin ashwin.devaraj@utexas.edu Byron C. Wallace Northeastern University b.wallace@northeastern.edu Iain J. Marshall King’s College London iain.marshall@kcl.ac.uk Junyi Jessy Li The University of Texas at Austin jessy@austin.utexas.edu Abstract We consider the problem of learning to sim- plify medical texts. This is important be- cause most reliable, up-to-date information in biomedicine is dense with jargon and thus practically inaccessible to the lay audience. Furthermore, manual simplification does not scale to the rapidly growing body of biomed- ical literature, motivating the need for auto- mated approaches. Unfortunately, there are no large-scale resources available for this task. In this work we introduce a new corpus of parallel texts in English comprising technical and lay summaries of all published evidence pertaining to different clinical topics. We then propose a new metric based on likeli- hood scores from a masked language model pretrained on scientific texts. We show that this automated measure better differentiates between technical and lay summaries than ex- isting heuristics. We introduce and evaluate baseline encoder-decoder Transformer models for simplification and propose a novel augmen- tation to these in which we explicitly penalize the decoder for producing ‘jargon’ terms; we find that this yields improvements over base- lines in terms of readability. 1 Introduction The need for accessible medical information has never been greater. A Pew Research survey of American’s online health habits in 2013 revealed that “one in three American adults have gone on- line to figure out a medical condition” (Fox and Duggan, 2013). Given the rise of medical misin- formation on the internet (Ioannidis et al., 2017), accessibility has become an increasingly urgent is- sue (World Health Organization, 2013; Armstrong and Naylor, 2019). However, sources that provide accurate and up-to-date information, including sci- entific papers and systematic reviews (Chalmers et al., 1995), are often effectively inaccessible to most readers because they are highly technical and laden with terminology (Damay et al., 2006). Technical abstract: Analysis showed a higher rate of weight gain in the high-volume feeds group: mean difference 6.20 g/kg/d (95% confidence interval 2.71 to 9.69). There was no increase in the risk of feed intolerance or necrotising enterocolitis with high-volume feeds, but 95% confidence intervals around these estimates were wide. Plain-language summary: Very low birth weight infants who receive more milk than standard volumes gain weight more quickly during their hospital stay. We found no evi- dence suggesting that giving infants high volumes of milk causes feeding or gut problems, but this finding is not certain. Table 1: Sample excerpts from a technical abstract (top) and corresponding plain-language summary (bottom) from the Cochrane Library. One potential solution to this problem is text sim- plification, i.e., editing documents such that they are accessible to a wider audience, while preserv- ing the key information that they contain. Although manual …

Content truncated to 500 words