Paragraph-level Simplification of Medical Texts

Abstract

We consider the problem of learning to simplify medical texts. This is important because 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 biomedical literature, motivating the need for automated 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 likelihood scores from a masked language model pretrained on scientific texts. We show that this automated measure better differentiates between technical and lay summaries than existing heuristics. We introduce and evaluate baseline encoder-decoder Transformer models for simplification and propose a novel augmentation to these in which we explicitly penalize the decoder for producing 'jargon' terms; we find that this yields improvements over baselines in terms of readability.

Paper Plain: Making Medical Research Papers Approachable to Healthcare Consumers with Natural Language Processing

“…Paper Plain leverages recent gains in natural language processing (NLP) for making medical information more understandable to the public, specifically healthcare consumers [31,104]. The research most salient to Paper Plain are automated term definition or replacement [102], plain language summarization [32], and consumer biomedical question answering [5]. In addition, we discuss here writing tools to encourage plain language [41], as the underlying techniques for powering such systems are similar to those leveraged by Paper Plain (e.g., generating plain language).…”

“…Paper Plain integrates these advancements in its implementation to show the promise of such methods in supporting Veyseh et al [102] presented a web-based system for acronym identification that works in the biomedical, scientific, and general domain and Murthy et al [75] explored how to define scientific terminology with terms recognizable to a specified reader. Devaraj et al [32] introduced a new dataset of healthcare consumer summaries for clinical topics and a trained model for simplifying medical text. Guo et al [42] used plain language summaries to train a model for generating summaries of biomedical text.…”

Evaluating Factuality in Text Simplification

“…Factuality (and the lack thereof) has been identified as critical in recent work in unsupservised simplification (Laban et al, 2021) and medical simplification (Devaraj et al, 2021). Guo et al (2018) incorporated textual entailment into their simplification task via an auxillary loss.…”

“…This may permit information accessibility to a wide range of audiences, e.g., non-native speakers (Yano et al, 1994), children (De Belder and Moens, 2010), as well as individuals with aphasia (Carroll et al, 1998) and dyslexia (Rello et al, 2013). Simplification may also help laypeople digest technical information that would otherwise be impenetrable (Damay et al, 2006;Devaraj et al, 2021).…”

“…An important but mostly overlooked aspect of automated simplification-especially in the conditional text generation regime-is whether outputs are faithful to the inputs that they are simplifying. Consider, for example, automatically simplifying medical texts (Devaraj et al, 2021): Presenting individuals with readable medical information that contains factual errors is probably worse than providing no such access at all. Recent work has acknowledged factuality and faithfulness as key issues to be addressed in other conditional generation tasks like summarization (Kryscinski et al, 2020a;Maynez et al, 2020;Pagnoni et al, 2021;Goyal and Durrett, 2021), yet so far little research has thoroughly studied the kinds of errors that simplification datasets and system outputs exhibit.…”