Biomedical concept recognition with error-aware negative-enhanced ranking framework

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Mention-agnostic biomedical concept recognition (MA-BCR) aims to identify concepts without explicit text spans.
• Existing MA-BCR methods primarily use generative or classification approaches.
• The effectiveness of ranking-based methods, specifically retrieve-rerank pipelines, for MA-BCR is underexplored.

Purpose of the Study:

• To systematically compare ranking-based approaches against generative and classification methods for MA-BCR.
• To investigate the optimal supervision strategies for training rankers in low-annotation settings.
• To introduce an improved ranking framework for MA-BCR.

Main Methods:

• A systematic comparison of ranking, generative, and classification paradigms was conducted.
• A two-stage retrieve-rerank architecture was established as a robust backbone for MA-BCR.
• An error-aware negative-enhanced ranking (ENR) framework was proposed, augmenting training with false positives.

Main Results:

• The retrieve-rerank architecture proved to be the most robust and scalable for MA-BCR.
• ENR significantly improved reranking performance by incorporating error-aware negative examples.
• ENR substantially outperformed existing methods on the MM-HPO and MM-GO datasets.

Conclusions:

• Ranking-based methods, particularly the retrieve-rerank paradigm, offer a powerful approach for MA-BCR.
• The proposed ENR framework enhances MA-BCR performance without increasing inference costs.
• ENR represents a significant advancement in MA-BCR, demonstrating superior results on benchmark datasets.