The emergence of online medical question-answer communities has helped to balance the supply of medical resources. However, the dramatic increase in the number of patients consulting online resources has resulted in a large number of repetitive medical questions, significantly reducing the efficiency of doctors in answering these questions. To improve the efficiency of online consultations, a large number of deep learning methods have been used for medical question-answer matching tasks. Medical question-answer matching involves identifying the best answer to a given question from a set of candidate answers. Previous studies have focused on representation-based and interaction-based question-answer pairs, with little attention paid to the effect of noise words on matching. Moreover, only local-level information was used for similarity modeling, ignoring the importance of global-level information. In this paper, we propose a dual-channel attention with global similarity (DCAG) framework to address the above issues in question-answer matching. The introduction of a self-attention mechanism assigns a different weight to each word in questions and answers, reducing the noise of "useless words" in sentences. After the text representations were obtained through the dual-channel attention model, a gating mechanism was introduced for global similarity modeling. The experimental results on the cMedQA vl.O dataset show that our framework significantly outperformed existing state-of-the-art models, especially those using pretrained BERT models for word embedding, improving the top-1 accuracy to 75.6%.
展开▼
