Document Retrieval for Biomedical Question Answering with Neural Sentence Matching

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Document retrieval (DR) is crucial for question-answering (QA) systems.
• Effective DR can provide answers even without advanced natural language processing (NLP) for extraction.
• Leveraging BioASQ datasets enables building specialized biomedical DR systems.

Purpose of the Study:

• To develop an effective biomedical document retrieval system for question-answering tasks.
• To utilize answer snippets from BioASQ training data for system development.
• To enhance document relevance scoring using multiple features.

Main Methods:

• Core approach: question-answer sentence matching neural network to calculate relevance scores.
• Auxiliary features: journal name and presence of semantic relations (subject-predicate-object triples).
• Reranking baseline scores using these features with adaptive random search and learning-to-rank methods.

Main Results:

• The developed system achieved 2nd place in the BioASQ 2018 Phase A (DR) of Task B (QA).
• Ablation experiments confirmed the significant contribution of the neural matching network component.
• The system effectively leverages semantic information and publication context for improved retrieval.

Conclusions:

• The proposed neural matching network is a significant component for effective biomedical document retrieval.
• Combining neural matching with auxiliary features like journal information and semantic relations enhances QA system performance.
• The approach demonstrates a successful strategy for building high-performing DR systems within the BioASQ framework.