Lexicon enhanced Chinese named entity recognition with pointer network

In recent time, lexicon-based LSTM and pre-training language models are combined to explore the Chinese Named Entity Recognition (NER) and achieve the current state-of-the-art (SOTA) performance on several Chinese benchmark datasets. However, existing lexicon-based models only conform lexicon features through shallow and randomly initialized coding layers and do not integrate them into the bottom layer of the pre-training language model to mine the deep lexicon knowledge. To address the above issue, we propose a novel BERT-based Enhanced Lexicon Adapter (BLA) model that fuses external lexicon feature into the pre-training language model BERT in-depth. Specifically, the external lexicon knowledge is integrated into the deep Transformer layers of BERT by the lexicon adapter mechanism. With the comparison of existing methods, our model achieves the genuine deep fusion of the lexicon knowledge and BERT representation, effectively obtaining entity boundaries and word information.Besides, given the value of high-level global semantic features in alleviating word ambiguity and segmenting precisely the entity boundary in Chinese NER, transforming the sequence labeling task into sequence generation task provides the new cogitation for extracting global semantic features. Therefore, we explore the strategies of local lexicon information’s fusion and global semantic features extraction for entity category labeling. Specifically, we utilize the sequence-to-sequence (Seq2Seq) framework with pointer network as the prominent model architecture, in which the pointing function implements a custom attention mechanism and models different interactions between the source text and the semantic embedding by the generated probability ppoint\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{point}$$\end{document}. Furthermore, the decoder with the pointer mechanism generates the target sequence autoregressively. Experiments on several different benchmark Chinese datasets indicate that the proposed model achieves remarkable improvement compared with the current lexicon-based methods, and the results significantly outperform the current SOTA models.