Semantic Matching via Interactive Attention Fusion

The patent introduces an attention mechanism as an intermediary component between the query and document processing streams. This attention mechanism computes interaction features by attending to both query and document embeddings, enabling the model to capture semantic relationships without sacrificing the efficient Siamese network architecture. The attention mechanism serves as a mediator that enriches the feature representation while maintaining computational efficiency.

The patent creates a composite feature representation by combining multiple feature types: original embeddings, self-attention features (capturing intra-sequence relationships), and interaction features (capturing inter-sequence relationships). This composite approach integrates multiple processing perspectives into a unified feature vector that preserves the speed benefits of the Siamese network while achieving higher matching accuracy through enriched feature composition.

The attention mechanism acts as an intermediary that enables interaction between the query and document sides without requiring complex bidirectional processing. By computing attention weights between query tokens and document tokens, the model captures interactive characteristics while maintaining the unidirectional processing flow that keeps the model structure relatively simple and computationally efficient.

The patent applies local quality by making different parts of the model have different functions: the embedding layers handle local token representations, self-attention layers handle local intra-sequence relationships, and the interaction attention layer handles specific query-document relationships. This differentiated local processing enables interactive characteristics without uniformly increasing the complexity of the entire model structure.