Improving Medical Question Summarization through Re-ranking
- Publisher:
- Institute of Electrical and Electronics Engineers (IEEE)
- Publication Type:
- Conference Proceeding
- Citation:
- 2025 International Joint Conference on Neural Networks (IJCNN), 2025, 00, pp. 1-8
- Issue Date:
- 2025-07-05
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Fine-tuning sequence-to-sequence (Seq2Seq) models applied on downstream datasets have gained remarkable success in the task of medical question summarization (MQS). However, Seq2Seq models suffer from a discrepancy between their objective function and evaluation metrics, whose objective function is based on local and token-level predictions whereas the evaluation metrics of MQS focus on overall similarity between the gold references and system predictions. To address the gap, it is crucial to consider multiple candidate summaries, assess their quality and re-rank them to obtain the optimal summary. In this paper, we propose to enhance MQS through re-ranking. Specifically, we introduce two re-rankers (SimGate and LLM) designed to perform reference-free evaluations on candidate summaries and select the best one. The former evaluates candidate summaries using text semantic similarity and a confidence gate, while the latter leverages the powerful comprehension capability of large language models to identify more preferable summaries. Experimental results demonstrate the effectiveness of the proposed re-ranking mechanisms. The SimGate re-ranker, compared with BART, exhibits a significant improvement in the ROUGE score on all datasets. Additionally, LLM re-rankers also demonstrate significant improvements on most datasets, underscoring the promising potential of large language models as re-rankers. The code and datasets are available at https://github.com/yrbobo/MQS-Reranker.
Please use this identifier to cite or link to this item: