UETQuintet at BioCreative IX - MedHopQA: Enhancing Biomedical QA with Selective Multi-hop Reasoning and Contextual Retrieval

TL;DR

The paper tackles biomedical QA by enabling selective multi-hop reasoning through a framework that distinguishes direct from sequential questions and applies decomposition only when needed. It combines multi-source retrieval (web search and Wikipedia), in-context learning, and answer normalization to ground responses, achieving an Exact Match of 0.84 and ranking second on BioCreative IX MedHopQA. Key contributions include a lightweight sequential-question classifier, a decomposition pipeline, and Wikipedia-based normalization that improves both EM and concept-level evaluation. The approach highlights the practical value of retrieval-augmented generation and prompt design for robust biomedical question answering in real-world tasks.

Abstract

Biomedical Question Answering systems play a critical role in processing complex medical queries, yet they often struggle with the intricate nature of medical data and the demand for multi-hop reasoning. In this paper, we propose a model designed to effectively address both direct and sequential questions. While sequential questions are decomposed into a chain of sub-questions to perform reasoning across a chain of steps, direct questions are processed directly to ensure efficiency and minimise processing overhead. Additionally, we leverage multi-source information retrieval and in-context learning to provide rich, relevant context for generating answers. We evaluated our model on the BioCreative IX - MedHopQA Shared Task datasets. Our approach achieves an Exact Match score of 0.84, ranking second on the current leaderboard. These results highlight the model's capability to meet the challenges of Biomedical Question Answering, offering a versatile solution for advancing medical research and practice.

Figures (3)

• Figure 1: Overview of the proposed model. (a) Question Decomposition: The input question is simplified and classified as either a direct or a sequential question. Sequential questions are further decomposed into sub-questions. For each question, a query and an initial anchor are extracted. (b) Answer Generation at hop $i^{th}$: The model retrieves relevant context and generates an answer using the sub-question, retrieved context and the anchor. The short answer output $a_i$ is used as the final predicted answer if $i$ is the last hop, otherwise, it is used as an anchor for the next hop $i+1$.
• Figure 2: System prompt for sequential question
• Figure 3: System prompt for answer generation