DGC-CRL: Dependency Graph Convolution based Contrastive Representation Learning for Chinese Medical Question Matching

Abstract

As one kind of domain-specific question answering (QA) systems, the medical QA systems require much more stability, fast system speed and response accuracy. Therefore, the retrieval based QA systems are more suitable, among which the deep semantic matching models become prevalent to be studied and they are playing the very important role on the quality of retrieval based medical QA systems. In this paper, we propose a two-stage solution (named with Dependency Graph Convolution based Contrastive Representation Learning) which includes a dependency graph convolution module to explicitly capture the semantic similarity between Chinese questions. At the first stage, we adopt the contrastive learning to further distinguish the similarity within the domain-specific corpus itself and learn discriminative textual representations. At the second stage, the down-streaming question matching task is benefited by using the newly-learned representations. In our experiments, we collect two Chinese medical datasets (CBLUE-STS and COVID-19) and the results can demonstrate that our proposed method is effective and general to different medical QA corpora. Also the ablation experiments indicate the proposed Dependency Graph Convolution module and contrastive learning method are both efficient.