A Diabetes Question-Answer Semantic Matching Model Integrating Multi-Granularity Information and Attention Mechanisms

With the rapid increase in the demand for medical question-and-answer systems in the field of diabetes, existing semantic matching models often overlook the extraction of multigranularity information and fail to effectively capture the interactive features both within and between the question text and the answer text. This limitation restricts the model's semantic understanding capabilities, making it challenging to meet the requirements for precise question answering. To address this, this paper proposes a diabetes question-and-answer semantic matching model (BMQA-KG) that integrates multi-granularity information and attention mechanisms, aiming to enhance the accuracy of question-answer matching. The model first employs BERT to encode character and word-level information within the question-answer pairs and utilizes BiLSTM to capture the semantic features at both the character and word levels of the question-answer pairs. Subsequently, it implements a soft alignment attention network and an absolute value subtraction attention network to extract interactive features both within the question-answer texts and between them. Additionally, it incorporates entity and relationship embeddings from a medical knowledge graph to enrich the fine-grained information within the question-answer texts. Experimental results demonstrate that the BMQA-KG model outperforms existing question-answer matching models on both the cMedQA2.0 and the self-constructed MQA datasets.