Research on English grammatical structure parsing method based on transformer and GCN

English grammatical structure parsing is a core task in natural language processing, playing a crucial role in machine translation, information extraction, and other applications. This paper proposes an innovative model, TG-Parser, that combines a Transformer and a graph convolutional network (GCN). It significantly improves the accuracy of English grammar parsing by combining the global semantic modeling capabilities of the Transformer and the structural relationship capture advantages of the GCN. First, a word-level context representation is obtained using a Transformer encoder; second, a fully connected grammar diagram is constructed, and a multi-layer GCN is used for structured feature learning; Finally, a dual-channel decoder is designed to handle dependency analysis and component analysis tasks respectively. Experiments on the Penn Treebank (PTB) and Universal Dependencies (UD) English datasets demonstrate that Labeled Attachment Score (LAS) achieves 96.2% Universal Dependencies—English Web Treebank (UD-EWT) and 95.8% Universal Dependencies—Georgetown University Multilayer Corpus (UD-GUM) accuracy in the dependency resolution task, surpassing performance of pure Transformer model by 2.1%; In component analysis task, the F1 Score reached 95.5% (PTB), exceeding the baseline model by 1.7%. At the same time, ablation experiments confirmed that the GCN module contributed significantly to long-distance dependency analysis (Unlabeled Attachment Score (UAS) increased by 1.8%). This study verifies the effectiveness of heterogeneous architecture fusion in grammar parsing and provides a new paradigm for modeling complex language structures.