Enhancing Chinese NER with Deep Contextual Representations and Attention Mechanisms

N amed Entity Recognition (NER) is a pivotal task in NLP, designed to discern and categorize entities such as individuals, geographical regions, and institutions within a given text2. While various methods have been proposed for English NER tasks, Chinese NER faces more significant challenges, including the lack of explicit entity indicators, high dependence on context, and the absence of clear word boundaries. For this purpose, this paper introduces a novel model architecture: ERNIE-BiGRU-Att-CRF. Initially, the ERNIE pre-trained model is utilized to capture rich character-level contextual embeddings, effectively grasping complex semantic nuances in Chinese text. Then, Bidirectional Gated Recurrent Units (BiGRU) are employed to extract global features, thereby enhancing the model's comprehension of semantic shifts. Building upon this, an attention mechanism is incorporated to dynamically assign different weights to crucial information, improving the model's capability to process complex entities and semantic links. Finally, Conditional Random Fields (CRF) are integrated as the decoding layer to ensure that the label sequence maintains global coherence and precision. The experimental results show that the proposed model achieves Fl scores of 85.38% and 79.13% on the benchmark datasets Weibo NER and EduNER, respectively, significantly outperforming existing methods. This indicates that the model proposed in this study yields good results. The framework not only improves the performance of Chinese NER but also provides an effective solution for handling long-distance dependencies and ambiguous entity boundaries.