基于注意力混合模型的中文医疗问答匹配

CHINESE MEDICAL QA MATCHING BASED ON ATTENTION HYBRID MODEL

将深度学习应用于医疗问答是目前最热门的话题之一,一般使用单一神经网络(CNN或RNN)获取句子间语义信息来提高匹配精度。但是该方法只能获取句子中一部分或单侧的信息,忽略了其他丰富而复杂的语义关系,从而导致匹配结果不够理想。针对此问题提出一种将CNN网络和BiGRU网络相结合并引入注意力机制的混合模型—ABiGRU-CNN,该模型能够有效地提取问答对中复杂语义信息。在cMedQA数据集上的实验表明,所提出的混合模型比现有方法具有更高的匹配精度。

It is one of the most popular topics to apply deep learning to medical question answering. Traditional methods generally use a single neural network(CNN or RNN) to obtain semantic information between sentences to improve the matching accuracy. However, this method can only obtain part or one side of the sentence information, ignoring other rich and complex semantic relations, which leads to the unsatisfactory matching results. To solve this problem, a hybrid model, ABiGRU-CNN, which combines CNN and BiGRU networks and introduces attention mechanism is proposed. This model could effectively extract complex semantic information from QA pairs. The experiments on cMedQA data set show that the proposed hybrid model has higher matching accuracy than the existing ones.