lasigeBioTM:A Lean Biomedical QA System Empowered by Structured Knowledge
Year: 2025
Type: conference-paper
Abstract: Biomedical Question Answering (BQA) presents unique challenges due to the vast and complex nature of biomedical literature and the highly specialized terminology involved. In the BioASQ13 challenge, we participated in all phases (A, A+, and B), proposing a hybrid system that integrates large language models (LLMs) with structured knowledge sources, namely ontologies and knowledge graphs. Our approach combines retrieval-augmented generation for document and snippet selection with graph-constrained decoding, which improves factual accuracy and supports reasoning grounded in biomedical knowledge. Its "lean" aspect arises from the use of a lightweight LLM and efficient techniques that minimize dependence on the extensive computational resources often associated with larger models. While our baseline system, lasigeBioTM, consistently performed well on exact answers across most batches, the integration of external information occasionally introduced bottlenecks or noise, highlighting how challenging it is to balance external knowledge and model fluency. Our results show that the system performance was closely tied to the quality of the retrieved snippets, with Phase B generally outperforming A+ due to better snippet quality. Our findings emphasize the value of high-quality retrieval and controlled knowledge integration for improving BQA performance. The code is publicly available at https://github.com/lasigeBioTM/BioASQ13_2025.
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7126997187