Med-EASi: Finely Annotated Dataset and Models for Controllable Simplification of Medical Texts

Automatic medical text simplification can assist providers with patient-friendly communication and make texts more accessible, thereby improving health literacy. But curating a quality corpus for this task requires the supervision of experts. In work, we present Med-EASi (Medical dataset Elaborative Abstractive Simplification), uniquely crowdsourced finely annotated supervised short texts. Its expert-layman-AI collaborative annotations facilitate controllability over by marking four kinds textual transformations: elaboration, replacement, deletion, insertion. To learn simplification, fine-tune T5-large different styles input-output combinations, leading to two control-free controllable versions model. We add types into using multi-angle training approach: position-aware, which uses in-place inputs outputs, position-agnostic, where model only knows contents be edited, but not their positions. Our results show that our fine-grained improve learning compared unannotated baseline. Furthermore, position-aware control enhances model's ability generate better than position-agnostic version. The data code are available at https://github.com/Chandrayee/CTRL-SIMP.