A clustering-based few-shot cross-domain named entity recognition method

Abstract

The application provides a clustering few-shot cross-domain named entity recognition method, and belongs to the technical field of natural language processing. The method comprises the following steps: obtaining a small amount of labeled target domain text data and sufficient source domain labeled text data; initializing a clustering center by using the target domain data, and clustering the source domain feature representation according to the clustering center, so that the source domain data is distributed in the semantic space and is aligned according to the potential category structure of the target domain; in the process of training the model, the category information of the source domain and the target domain is fused, and a multi-granularity inter-class and intra-class balance loss function is introduced, so that the category distinguishing ability and feature compactness of the model are optimized; and finally, a named entity recognition model with good generalization performance on the target domain is obtained. By guiding the clustering initialization and feature alignment process by using the target domain data, the application effectively alleviates the performance decline problem caused by the field difference and the inconsistent category granularity in the cross-domain named entity recognition task, is especially suitable for the few-shot scene with scarce labeled data of the target domain, and significantly improves the recognition accuracy and robustness of the model for cross-domain named entities.

Images