Bilingual dictionaries, as a fundamental language resource, play an important role in cross-language communication, providing key support for vocabulary mapping between different languages and serving as a basic tool for achieving language understanding and translation. To address the shortcomings of traditional methods, the corpus-based bilingual dictionary extraction method emerged. This study proposes a new deep learning model and presents a heterogeneous graph construction method that integrates parallel corpora and comparable corpora. It makes full use of the processing ability of GNN for structural information and breaks through the strong dependence of traditional methods on parallel corpora. By embedding similarity to construct entity-level translation edges, combined with GAT attention to dynamically adjust edge weights, and through the cross-language contrast attention mechanism to force translation nodes to align in the representation space, a full-process closed-loop design from heterogeneous data fusion, deep semantic modeling to generation optimization is achieved for the first time. In low-resource scenarios, the accuracy is significantly improved compared to the baseline model.