基于多特征信息融合自注意机制的中文命名实体识别方法

A Chinese-Named Entity Recognition Method Based on Multi-FeatureInformation Fusion and a Self-Attention Mechanism

工业领域数据由于其非结构化、领域特定性和数据稀缺性等特点,传统的中文命名实体识别技术在工业领域的应用并不理想。本文以汽车产业数据为依托,提出一种将标签语义与字形拼音信息相融合的自注意网络算法,结合字符级和标签级特征进行多维特征提取。模型引入自注意机制获得文本长距离依赖关系,将分词特征整合到字符级,并结合标签语义特征的上下文进行预测,提高了字符词边界的识别性能。在一定程度上解决了词边界分割歧义及短语组合上下文依赖问题。本文方法在MSRA和Weibo数据集及自构建工业维修文档数据集上进行了实验,结果表明所提方法能够提高实体识别准确性,并在工业领域汽车零配件数据集上实现了工业场景化应用。

Due to the inherent characteristics of industrial domain data,such as its unstructured format,domain-specific nature,and the scarcity of available data,the application of traditional Chinese-Named Entity Recognition(NER)techniques in industrial contexts often leads to suboptimal results.In response,this paper introduces a novel self-attention network algorithm,leveraging automotive industry data,that synergizes label semantics with glyph and phonetic elements.This model employs a self-attention mechanism to discern long-distance textual dependencies and integrates these with word segmentation at the character level.The model notably enhances character and word boundary recognition by fusing these elements with the context of label semantics.This approach effectively mitigates the common ambiguities in word boundary segmentation and the complexities associated with phrase contextual dependencies.Extensive testing on the MSRA and Weibo datasets,and a bespoke industrial maintenance docu⁃ment dataset demonstrates the methods efficacy.The results reveal a significant improvement in entity recognition accuracy,with specific enhancements in industrial scenario applications,particularly in the automotive parts sector datasets.