The article presents a novel task, Text2MDT, which aims to automatically extract medical decision trees from medical texts such as guidelines and textbooks. Two methods for the task are investigated: an end-to-end framework based on a GPT style large language model, and a pipeline framework which breaks the task into three subtasks. Experiments show that the end-to-end method shows promising results and outperforms the pipeline methods. The Text2MDT dataset used is open-sourced.
Publication date: 5 Jan 2024
Project Page: https://tianchi.aliyun.com/dataset/95414, https://github.com/michael-wzhu/text2dt
Paper: https://arxiv.org/pdf/2401.02034
