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.
Introduction
The accessibility of medical literature is a significant barrier to health literacy. Most high-quality evidence, such as systematic reviews, is written in dense, technical language that is inaccessible to lay audiences. While automated text simplification offers a potential solution, existing research often suffers from domain mismatch and a focus on sentence-level transformations that fail to capture the complexity of medical discourse. This paper addresses these challenges by introducing a new paragraph-level dataset and a novel training objective designed to suppress technical jargon.
The Cochrane Dataset
To facilitate research in this area, the authors curated a new corpus from the Cochrane Library, a gold-standard source for medical systematic reviews. This dataset pairs technical abstracts with Plain-Language Summaries (PLS) authored by medical experts. After rigorous filtering and alignment, the resulting dataset consists of 4,459 high-quality paragraph pairs. Unlike previous sentence-aligned datasets, this section-aligned corpus allows models to learn more sophisticated, abstractive transformations necessary for medical simplification.
Methods: Jargon Suppression and Evaluation
The proposed approach utilizes a BART (Bidirectional and Auto-Regressive Transformer) model as a baseline. To enhance its simplification capabilities, the authors introduce two primary innovations:
- Unlikelihood Training: Beyond standard maximum likelihood estimation, the authors incorporate a secondary objective that explicitly penalizes the generation of "jargon" terms. These terms are identified using a logistic regression classifier trained to distinguish between technical and lay vocabulary.
- SciBERT-based Evaluation: Recognizing the limitations of traditional readability metrics, the authors propose a new evaluation framework based on SciBERT. By analyzing the predictability of tokens within a text using a model pretrained on scientific literature, they can more accurately quantify the "technicality" of a document.
Key Results and Insights
A central finding of this work is that traditional readability metrics, such as the Flesch-Kincaid Grade Level, are often insufficient for medical texts. These metrics rely on surface-level features like word length, which do not fully capture the complexity of medical jargon. In contrast, the SciBERT-based metric demonstrated a superior ability to distinguish between technical abstracts and lay summaries.
Models trained with the jargon penalty (UL-Cochrane) showed significant improvements in readability. Specifically, the UL-Cochrane model achieved a Flesch-Kincaid score of 11.97, outperforming both the standard BART baseline (13.44) and the human-written PLS (13.11). By penalizing statistical markers and technical terms, the model is incentivized to produce more abstractive and accessible summaries.
Conclusion and Future Work
While the proposed methods significantly improve readability, challenges remain. The models occasionally exhibit "hallucinations," such as inserting generic disclaimers not present in the source text. Furthermore, there is an inherent trade-off between simplification and the preservation of clinical detail. Future work should focus on ensuring clinical accuracy and exploring ways to balance information density with accessibility.
Summary updated