Representation Learning for Traditional Chinese Medicine Texts Using Graph Convolutional Networks with Integrated Syntactic and Semantic Features

The anomalous syntactic structures and linguistic conciseness inherent in traditional Chinese medical texts give rise to inadequate representation of inter-word dependency relationships and semantic information gaps during representation learning for this specialized corpus. To address these challenges, we propose a novel representation learning model, termed WGCN. We apply dependency syntactic parsing to address the challenge of obscured syntactic relationships between words. The resulting dependency trees are subsequently processed by an enhanced Graph Convolutional Network to extract features and generate word embeddings. To compensate for potential semantic gaps, sensum information is incorporated to construct a sememe-relation graph that captures lexical semantic relationships. The word embeddings produced by the enhanced module serve as the initial node features for the constructed graph. Finally, Graph Convolutional Network are applied to perform deep semantic feature extraction, producing word vectors that integrate both syntactic and semantic information tailored for traditional Chinese medical texts. Comprehensive evaluations on traditional Chinese medical corpora were carried out across both intrinsic and extrinsic tasks. The WGCN model achieved superior performance, with improvements of 0.62% and 0.97% over competing models, respectively. These results validate the model's enhanced representation learning capability.