Retrieval-Augmented Biomedical Question Answering with Weak-Question Recovery and Neural Reranking for BioASQ Task 14b

Research Question and Approach

This paper addresses the persistent challenge of retrieval quality in biomedical question answering (QA). In the context of the BioASQ Task 14b benchmark, the authors aim to overcome the limitations of standard retrieval systems, which often struggle with domain-specific terminology, synonym variability, and complex multi-hop reasoning. The proposed pipeline, developed by the DS@GT ARC team, focuses on enhancing evidence quality before it reaches the answer generation stage.

The approach centers on a multi-stage retrieval framework. First, it employs multi-source query expansion, utilizing training-memory, biomedical synonyms, and relationship-aware augmentation. Second, it uses a fine-tuned MiniLM-based neural reranker to score document relevance. A key innovation is the conditional weak-question recovery module, which identifies poorly performing queries—based on snippet quality and lexical overlap—and triggers a secondary, more intensive retrieval process. Finally, the system applies a pruning stage to remove redundant or irrelevant snippets, ensuring that the evidence provided to the OpenBioLLM generative model is both concise and high-quality.

The researchers demonstrate that their integrated pipeline significantly improves retrieval robustness. Experimental results on BioASQ evaluation batches show that the combination of weak-question recovery and post-retrieval pruning leads to a substantial increase in performance. Specifically, the final pipeline achieved a MAP@10 of 0.0955, representing a 28% relative improvement over the baseline retrieval system. The authors also highlight that snippet pruning is critical; their analysis shows that evidence coverage reaches diminishing returns after approximately two snippets, justifying their strategy of prioritizing quality over quantity to reduce noise for the LLM.

Biomedical QA systems are essential for helping researchers and clinicians navigate the rapidly expanding volume of medical literature. By focusing on the retrieval bottleneck—where noisy or incomplete evidence often leads to hallucinated or inconsistent answers—this work provides a practical blueprint for building more reliable RAG systems. The use of Activation-aware Weight Quantization (AWQ) to deploy a large 70B parameter model on a single GPU further demonstrates how high-performance reasoning can be achieved under constrained hardware resources, making advanced biomedical AI more accessible.