Rectifying and Discriminating Hard Negatives for Biomedical Retrieval Question Answering.

Area of Science:

• Biomedical Natural Language Processing
• Information Retrieval
• Machine Learning

Background:

• Bi-encoders are efficient for biomedical retrieval question answering (ReQA) but struggle with fine-grained interactions.
• Data scarcity in the biomedical field exacerbates training limitations for bi-encoders.
• Hard in-batch negative sampling improves training but introduces false negatives and embedding quality issues.

Purpose of the Study:

• To address challenges in training bi-encoders with hard negative samples in ReQA.
• To propose a novel framework, Rectifying and Discriminating Hard negatives (RigHt), to enhance bi-encoder performance.
• To improve the accuracy and robustness of biomedical ReQA systems.

Main Methods:

• Developed the RigHt framework to rectify false negative labels using cross-encoder interaction.
• Enhanced bi-encoders' ability to discriminate hard negative samples via refined disentangled sentence embeddings.
• Evaluated the framework on five diverse biomedical datasets.

Main Results:

• RigHt effectively mitigates the impact of false negatives during model training.
• The framework significantly improves the bi-encoder's capacity to distinguish hard negative samples.
• Experimental results demonstrate substantial performance gains across multiple datasets.

Conclusions:

• The RigHt framework offers a robust solution for training bi-encoders with hard negative samples in ReQA.
• This approach enhances the discriminative power of bi-encoders in the challenging biomedical domain.
• RigHt represents a significant advancement in improving the efficiency and accuracy of biomedical information retrieval.