为了提升 GlobaiPointer 方法的实体边界区分性能,提出一种融合词汇信息与 GlobalPointer 的实体识别方法.对SoftLexicon提取的词汇特征与字符相结合,采用BiLSTM网络与RoPE编码捕捉时序与相对位置信息构建全面特征,通过实体矩阵实现实体识别.对多个数据集进行试验,本研究提出的模型相较于其他基线模型,精确率、召回率、F1 均有一定的提升,Weibo数据集中F1 达到 71.33%、CMeEE数据集中F1 达到 63.45%,表明本研究提出的模型架构能够进一步扩充语义表征,增强识别性能.
In order to improve the entity boundary differentiation performance of GlobalPointer,an entity recognition method integrating lexicon information and globalpointer was proposed to enhance the recognition performance.For characters,softlexcion was used to extract vocabulary features and combine them with characters.BiLSTM network and RoPE code were used to capture timing and relative position information to construct comprehensive features.Entity recognition was realized through entity matrix.Experiments were carried out on multiple datasets.Compared with other baseline models,the model had made some progress in the metrics of precision,recall and F1.The F1 in Weibo dataset had reached 71.33%,and the F1 in CMeEE dataset had reached 63.45%.It indicated that the model architecture could further expand semantic representation and enhance recognition performance.