基于增强跨度信息表示的中文命名实体识别OA

Named entity recognition(NER)is one of the fundamental tasks in natural language processing.Due to the ina-dequate exploitation of semantic information within text spans,most previous works on Chinese NER suffer from limited feature extraction,which consequently leads to the underperformance of NER models.Furthermore,the interrelationships between adjacent spans are not given due consideration.To address these issues,a Chinese named entity recognition app-roach based on enhanced span information representations is proposed.The method employs two core modules:the span filter that evaluates entities and non-entities by computing scores from feature representations of start and end tokens,with position information embedded;and the span classifier that fuses boundary and internal information to build enh-anced span representations for span scoring as per entity types,assisted by a single 2D convolutional layer for cross-span interaction to rectify classification scores.For each span,the sum of scores output by the two modules determines the pre-diction.The F1 scores of the proposed model reach 96.71％,96.15％,81.88％and 75.28％respectively on four Chinese named entity recognition tasks:Resume,MSRA,CLUENER2020 and CMeEE-V2.The results of ablation experiments demonstrate the effectiveness of each component in this method.