TSS MLGNN: A Multi Level Graph Neural Network Approach for Text Semantic Similarity Computation

Abstract

In response to the challenges of short text noise and complex semantic associations in online social networks and media, a new method for calculating text semantic similarity is proposed. The purpose of this method is to overcome the limitations of traditional methods in identifying deep semantic relationships, thereby improving accuracy and efficiency. This method proposes a multi-level graph representation learning method based on graph neural network, and establishes a text semantic similarity calculation model based on graph neural network. The graph neural network is used to dynamically learn the complex relationship between nodes, and the multi view semantic features are extracted in parallel with the multi head attention mechanism. Finally, the hierarchical aggregation strategy is used to generate a high-order graph embedding representation that integrates local details and global semantics. Experiments were conducted on two common benchmark datasets, the semantic text similarity benchmark (STS-B, including subsets a-d) and the quora question pairs dataset (QQPD). Performance was evaluated using multiple metrics including accuracy, F1 score, Pearson correlation, calculation time, and memory consumption. Compared to baseline models such as Roberta, Siamese LSTM, and Ernie, the proposed model achieved the highest prediction accuracy. It had F1 scores of 94.21% and 91.38%, respectively, as well as a Pearson correlation coefficient of 0.85. The calculation time and memory consumption were reduced by 26.21% and 13.55% on average. These results demonstrate the effectiveness and robustness of the method in capturing fine-grained and long-distance semantic dependencies, particularly in noisy and informal social media environments. The method is highly applicable to real-time information review and rumor detection. It provides a more accurate computing tool for semantic analysis tasks in online social networks and media scenarios, and has a wide range of practical value in applications such as real-time information review, rumor detection and recommendation systems.