Document similarity detection based on multi-feature semantic fusion and concept graph

Document Similarity Detection (DSD) is important for checking about plagiarism and duplicate project approval. Abundant word-based detection methods are not enough for actual requirements. The semantic level DSD has recently become a challenging direction because semantic information needs mine. Firstly, we propose a Contextual Multi-feature Semantic Fusion (CMSF) module to purify keyword recognition, and then propose a Concept Graph (CG) structure for document entity. CMSF employs multi-feature enhanced semantics to represent entity, and CG leverages the text interaction graph and text semantic representation vector as initial features, and then Graph Convolutional Network (GCN) uses the features to match document. Extensive experiments based on the real dataset show that CMSF-GCN may converge quickly and get a high accuracy rate compared with recent methods about natural language matching.