FedINER: Chinese Industrial Named Entity Recognition with Federated Learning

Named entity recognition (NER) is foundational in constructing industrial knowledge graphs and is an essential component in the automation of knowledge within the industry. However, due to the particularities of the industrial field, its data involves issues such as business competition and user privacy. Traditional machine learning algorithms concentrate all data on a single machine for training, which may lead to privacy breaches. Federated learning, a distributed machine learning architecture, updates global model parameters through a central server’s aggregation algorithm without exposing private data. Therefore, this paper proposes a privacy-preserving framework for Chinese named entity recognition in the industrial domain, FedINER. The local client model employs a grid-tagging-based approach, which expands upon the character pair tagging system of SOTA models, emphasizing entity boundary information, internal character pair relationships, and the types of relationships. Besides, a bi-axial attention mechanism is utilized to better capture the relevance of relationships within the character pair representation, further optimizing model performance. Then, through federated training, each client benefits from the collective optimization of the global model, thereby improving the performance and accuracy of the model. Experiments on two distinct industrial datasets, conducted under the premise of protecting data privacy and security, have validated the effectiveness of this approach.