EDEL: enhancing dense retrievers for curation of biomedical knowledge bases.

Area of Science:

• Biomedical Informatics
• Information Retrieval
• Machine Learning

Background:

• Accurate literature retrieval is crucial for building high-quality biomedical knowledge bases.
• Traditional methods like BM25 are being surpassed by dense retrieval models.
• Current training methods for dense retrieval struggle with partially available evidence and limited use of knowledge base structures.

Purpose of the Study:

• To develop an improved dense retrieval model for biomedical knowledge base curation.
• To address limitations in training data by incorporating partially grounded evidence and structured negative sampling.
• To accelerate the process of finding relevant papers for annotation by curators.

Main Methods:

• Introduced EDEL, a novel dense bi-encoder for biomedical literature retrieval.
• Developed a new loss function utilizing graded relevance scores instead of binary labels.
• Implemented a structured sampling strategy for training with diverse and challenging negative examples.

Main Results:

• EDEL demonstrated superior performance in precision oncology and post-translational modification curation settings.
• Achieved improvements of 1.5 and 3.4 percentage points in NDCG@10 compared to state-of-the-art models.
• Ablation studies confirmed the effectiveness of the proposed loss function and sampling strategy.

Conclusions:

• EDEL significantly enhances literature retrieval for biomedical knowledge base curation.
• The model facilitates faster and more accurate paper annotation for curators.
• EDEL represents a substantial advancement in applying dense retrieval to biomedical informatics.