中文命名實體識別（NER）作為自然語言理解領域的一項基礎任務， 在從非結構化文本中萃取結構化知識方面扮演著關鍵角色。然而， 其效能常受中文語言固有的歧義性所限制，例如缺乏明確的詞彙邊 界及大寫標記，這對其識別準確性構成重大挑戰。為應對此挑戰， 本研究提出了一種具備魯棒性(robustness)及泛化能力的中文命名 實體識別框架。此框架運用預訓練語言模型以生成深度的上下文嵌 入表徵。這些嵌入表徵再整合進上下文感知模組，此感知模組旨在 增強位置理解與消除歧義性。此外，該框架亦整合一種對抗式訓練 技術以提升模型魯棒性，並結合條件隨機場（CRF）層來確保最終輸 出的結構一致性。為驗證此框架的有效性，我們在來自社交媒體與 醫療保健領域的資料集上進行了全面的評估。實驗結果證明，我們 所提出的框架在性能表現上超越了當前最先進的模型。 ;Chinese Named Entity Recognition (NER) serves as a cornerstone to transform unstructured text into structured knowledge. However, its efficacy is frequently constrained by inherent linguistic ambiguities in the Chinese language, such as the absence of explicit word boundaries and capitalization, which poses significant chal- lenges to recognition accuracy. In response to this, we introduce a robust and gener- alizable framework for Chinese NER. The proposed framework utilizes a pre-trained language model to generate deep contextual embeddings. These embeddings are integrated with a contextual awareness module to enhance positional understand- ing and resolve ambiguity. Furthermore, the framework incorporates an adversarial training technique to improve model robustness and a conditional random field layer to ensure structural coherence of the final output. To validate the effectiveness of the framework, we conducted comprehensive evaluations on diverse datasets from the social media and healthcare domains. The experimental results reveal that our proposed framework outperforms existing leading models in Chinese NER.
