How to Extract and Interact? Nested Siamese Text Matching with Interaction and Extraction

Jianxiang Zang, Hui Liu

Abstract: Text matching is a core problem in Natural Language Processing, and one popular approach is the Siamese model. To enhance the performance of the Siamese matching model, researchers have proposed diverse network blocks, which can be divided into Information Interaction Blocks and Feature Extraction Blocks based on their functions. Traditional approaches, employing a pipeline-based arrangement for these blocks, encounter limitations when it comes to capturing focused semantics via local reasoning. To better systematize these two types of blocks, we propose a nested Siamese text matching framework, with the introduction of the Interaction Extraction Module (IEM) and the Function Amplification Module (FAM). Specifically, the IEM acts as a Feature Extraction or Information Interaction Block, nested within both the FAM and the model structure itself. IEM assumes the responsibility for global matching within the model while FAM leverages the Late Attention and nested IEMs to sharpen local matching and reasoning capabilities. Through our experiments, the performance of this nested Siamese model has proven to be superior to that of other pipeline-based Siamese matching models.