A Sentence Semantic Matching Model Based on Cross-Attention Mechanism

Most of the current deep learning based sentence semantic matching uses Siamese network to extract semantic features and then uses simple attention mechanism for interaction, which is effective for extracting coarse sentence semantics, but it is difficult to extract deep interaction matching information, and it is difficult to distinguish sentences with similar structures but different semantics. Therefore, a sentence semantic matching model based on cross-attention mechanism is proposed to build an interaction module with cross-attention mechanism to extract interaction matching information of sentences and learn richer interaction features of sentences; meanwhile, a Roformer-ft pretraining model is used to embed and encode sentences so that the model can learn fine-grained semantic differences of sentence pairs more effectively. It is demonstrated experimentally that the model in this paper exhibits better matching performance.