The article presents SPEER (Sentence-level Planning via Embedded Entity Retrieval), a method to improve the generation of long clinical summaries, a task often undertaken when a patient is discharged from a hospital. This task is time-consuming and complex due to the number of unique clinical concepts covered during the patient’s admission. The authors fine-tuned open-source Large Language Models (LLMs) and found that they produced incomplete and unfaithful summaries. To remedy this, they trained a smaller model to predict salient entities which would then guide the LLM. The method showed gains in both coverage and faithfulness metrics over non-guided and guided baselines.
Publication date: 4 Jan 2024
Project Page: https://arxiv.org/abs/2401.02369v1
Paper: https://arxiv.org/pdf/2401.02369
