Deep Learning for Chinese Toponym Recognition: A BiLSTM-CRF Based Approach

Chinese toponym recognition faces significant challenges due to linguistic complexity and diverse naming patterns. While existing studies rely on rule-based or statistical models, these approaches suffer from manual feature engineering and poor adaptability to complex or nested toponyms. This study proposes an innovative BiLSTM-CRF model enhanced with domain-specific data processing techniques to address these limitations. We constructed a large-scale annotated corpus by integrating Encyclopedia of China: Chinese Geography and Wikipedia. To capture polysemous features, we introduced a 20-dimensional disambiguation matrix combined with 100-dimensional pre-trained word vectors, forming a 120-dimensional representation. The BiLSTM network extracts bidirectional contextual features, while the CRF layer optimizes label sequences. Experiments on three corpora demonstrate that our model achieves state-of-the-art performance, with precision, recall, and F1 scores of 95.09%, 93.17%, and 94.12% on the domain-specific corpus, outperforming traditional CRF by 5.27% in F1. The integration of pre-trained embeddings and disambiguation mechanisms significantly enhances adaptability to diverse textual contexts. This work provides a robust framework for Chinese toponym recognition, emphasizing scalable data processing and hybrid deep learning architectures.