Semantic Matching Model based on Layer-Wise Attention Pooling Network and Dynamic Feature Fusion Mechanism

In semantic matching tasks, it is challenging to construct high-quality text vector representations and dynamically adjust the feature weights of keywords and abstract concepts in the text so that they are incorporated into the matching model in the optimal ratio. In this paper, we propose a novel BERT-based semantic matching model that aims to solve this problem. Firstly, a Layer-Wise Attention Pooling Network is introduced in the output of BERT, which can capture more important features by paying attention to and fusing the output of each layer through the attention mechanism, thus improving the text representation. Second, a Dynamic Feature Fusion Mechanism is proposed for dynamically adjusting the weights of keywords and abstract concepts in semantic matching to fully fuse the feature information of both. To evaluate the performance of the proposed model, experiments are conducted on MRPC, QQP, and Medical-SM datasets. The results show that compared to the existing models, the method proposed in this paper achieves significant accuracy improvement in the semantic matching task.