Research on Text Similarity Algorithm for Fusion of Multiple Features

Traditional text similarity calculation methods often have shortcomings when dealing with complex semantic and contextual relationships. To address this issue, a new model combining interactive and representational structures has been proposed. The embedding layer of this model uses Robustly Optimized BERT Approach(RoBERTa) to extract word embedding, and uses the self-attention calculation weights in RoBERTa as the self-attention map. It integrates knowledge graph, syntactic dependency graph, and word co-occurrence graph, and generates comprehensive graph features through the graph feature extraction layer. Then, the extracted graph feature information is fused with the word embedding through the Relational Graph Convolutional Network (rGCN) graph convolutional layer. The Bidirectional Gated Recurrent Unit(BiGRU) encoding layer is input to reinforce the semantic information contained in the context, and the final sentence vector feature after information interaction is obtained through attention mechanism. The experimental results show that this method effectively improves the accuracy of the model when dealing with Chinese text similarity tasks, with an accuracy of 89.48 and an F1-score of 89.47.