Enhancing Biomedical ReQA With Adversarial Hard In-Batch Negative Samples.

Area of Science:

• Biomedical Natural Language Processing
• Information Retrieval
• Machine Learning

Background:

• Biomedical question answering (QA) is crucial for processing health information.
• Retrieval question answering (ReQA) efficiently finds answers from candidate sets.
• Biomedical ReQA (BioReQA) applies ReQA to the medical domain, often using dual-encoder models and dense retrieval methods.

Purpose of the Study:

• To address the limitations of using easy negative samples in BioReQA training.
• To propose an effective method for constructing hard in-batch negative samples.
• To enhance BioReQA performance, especially in low-resource settings.

Main Methods:

• Developed the Iterative Linear Assignment Grouping (ILAG) algorithm, inspired by the linear assignment problem, to generate hard in-batch negative samples.
• Employed adversarial training to further increase the difficulty of training batches in low-resource scenarios.
• Evaluated the proposed method on biomedical retrieval question answering tasks.

Main Results:

• The proposed ILAG algorithm effectively constructs hard in-batch negative samples.
• Adversarial training further boosts performance in low-resource BioReQA.
• Experimental results demonstrate the method's significant potential for improving BioReQA.

Conclusions:

• The ILAG algorithm offers a promising approach for generating effective hard negative samples in BioReQA.
• The combination of ILAG and adversarial training enhances BioReQA performance, particularly under data constraints.
• This research contributes to advancing efficient and accurate biomedical information retrieval.