Text Semantic Matching Model Based on Adversarial Training and Multi-angle Features

Text semantic matching is a fundamental task in natural language processing, widely applied in areas such as information retrieval, question answering systems, and recommendation systems. Traditional matching methods primarily focus on shallow features, but recent advances in deep learning have significantly propelled the development of text matching techniques. In light of the challenges faced by existing models, this paper proposes a text semantic matching model based on adversarial training and multi-angle features (MATM). This model adopts a siamese network framework, integrating various neural network structures. It leverages the WoBERT pre-trained model to extract deep contextual semantic information and combines the strengths of BiLSTM and CNN in feature extraction. Additionally, the model introduces an attention mechanism to merge multi-angle semantic features and employs adversarial training to generate adversarial samples, enhancing the model's generalization ability. Experiments on two Chinese text semantic matching datasets, BQ and BDZD, demonstrate that the model achieves accuracy and F1-Score of 85.93% and 85.77% on BQ, 85.38% and 86.72% on BDZD, respectively, showcasing substantial competitiveness.