摘要： 现有命名实体识别模型在字嵌入过程中多采用字符向量、字向量等不同单词表示向量的拼接或累加方式提取信息，未考虑不同单词表示特征之间的相互依赖关系，导致单词内部特征信息获取不足。提出一种基于交互式特征融合的嵌套命名实体识别模型，通过交互的方式构建不同特征之间的通信桥梁，以捕获多特征之间的依赖关系。采用交互机制得到包含不同单词表示信息的字嵌入向量，基于双向长短时记忆网络提取单词的表示特征，并对不同单词的表示特征进行交互，捕获特征之间的相互依赖关系。为进一步提取序列特征的上下文信息，采用基于特征交互的多头注意力机制捕获句子上下文的依赖关系。在此基础上，采用二元序列标记法过滤非实体区域，得到粗粒度候选区间，并对其进行细粒度划分以判断实体类别。实验结果表明，该模型的召回率和F1值为72.4%和71.2%，相比现有的嵌套命名实体识别模型，F1值平均提高了1.72%。

关键词: 嵌套命名实体识别, 双向长短时记忆网络, 特征交互, 多头注意力, 候选区间

Abstract: During word embedding in existing Named Entity Recognition(NER) processes, information is extracted by splicing or accumulating different word representation vectors such as character and word vectors without considering the interdependence between different word representation features.This results in insufficient acquisition of internal feature information of words.Accordingly, a nested NER model based on interactive feature fusion is proposed in this study.A communication bridge between different features is constructed using an interactive approach to capture the dependency between multiple features.An interaction mechanism is used to obtain the word embedding vectors containing different word representation information, and a Bidirectional Long Short-Term Memory(BiLSTM) network is used to extract word representation features.The proposed model enables the representation features of different words to interact so that the interdependence between features can be captured.To further extract the contextual information of sequence features, a feature interaction-based multi-head attention mechanism is used to capture the dependency of sentence contexts.A binary sequence labeling method is then used to filter the non-entity regions, and coarse granularity candidate intervals are obtained that are then divided into fine granularity candidate intervals.Entity categories are then determined.Experimental results show that the recall rate and F1 value of the proposed model are 72.4% and 71.2%, respectively, and the F1 value increases by 1.72% on average as compared with the existing nested NER model.

Key words: nested Named Entity Recognition(NER), Bidirection Long Short-Term Memory(BiLSTM) network, feature interaction, multi-head attention, candidate intervals

中图分类号: 

• TP391