LSSL: Label Semantic and Scoring Label for Fine-grained Chinese Named Entity Recognition

Chinese Named Entity Recognition (CNER) is gaining prominence as Natural Language Processing (NLP) expands across various disciplines. The current methods mainly exploit additional information and ignore the interrelationships between hierarchical labels. Therefore, the Label Semantic and Scoring Label (LSSL) method is proposed for solving Fine-Grained Named Entity Recognition (FG-NER) problems based on soft labels. The LSSL consists of the character embedding module, label feature extraction module, and scoring label method. First, the character embedding module is achieved by the pre-training language model to extract not only semantic characteristics of text data but also the semantic roles and relationships of words in sentences. Then, label feature extraction module converts the label in natural language form with a labeling annotation scheme: BMES scheme into label representation, which is commonly used to identify and classify entities in text, such as people, place names, organization names, and so on. In BMES, B indicates the start position of the entity. M indicates the middle part of the entity. E indicates the end position of the entity. And S indicates an isolated entity. Finally, the scoring label method based on label hierarchical structures is designed to represent the relevance between flattened hierarchical labels. Comparative experiments are conducted on the Government Proposal Named Entity (GPNE) dataset, and the commonly used public datasets with Resume and Weibo. In the GPNE dataset, compared to the state-of-the-art models the LSSL achieved an average improvement of 3.88% in the micro-f1 metric. In Resume and Weibo datasets, the LSSL exceeds 1.08% on average, compared to Bert-Flat. The LSSL model can well extract the interrelationships information between hierarchical tags, and can absorb additional semantic details about tags to enhance Fine-Grained Chinese Named Entity Recognition (FG-CNER) performance.