Combining Multi-granularity Text Semantics with Graph Relational Semantics for Question Retrieval in CQA

Question retrieval aims to retrieve historical question-answer pairs that are semantically similar or related to newly posted questions. Existing methods rely on measuring the textual similarity between the asked question and the solved question, but suffer from insufficient semantic mining and inaccurate matching feature extraction. To address these issues, we propose a novel model that considers fine-grained word-level similarities and graph-based semantic relationships between questions, as well as potential sequence correlations between questions and answers. Specifically, a tag-enhanced multi-granularity matching strategy is designed to learn the semantic similarity between questions, and a BERT-based correlation mining method is adopted to explore the relevance between questions and answers. In addition, we construct a homogeneous question network based on the pointing relationships between question knowledge units and learn the relational semantics of question nodes through an auxiliary information-enhanced skip-gram algorithm. Evaluation results on two community datasets show that our proposed model significantly improves retrieval accuracy and efficiency compared to state-of-the-art methods.