Research On Chinese Named Entity Recognition Algorithm Based On Fusion Of Multiple Semantic Features

Chinese named entity recognition technology aims to identify entities and their categories from raw text,thereby improving the efficiency of information extraction.It is one of the most important and fundamental tasks in natural language processing and plays an important role in many downstream tasks.Currently,Chinese named entity recognition algorithms based on deep learning have achieved promising results.However,due to the particularity and complexity of the Chinese language structure,Chinese named entity recognition algorithms still face difficulties such as entity boundaries uncertainty and word ambiguity.Aiming at the difficulties and challenges faced by current Chinese named entity recognition,this paper enhances the input by acquiring multiple external knowledge and combines them through feature fusion networks to improve the performance of Chinese named entity recognition.The specific research content is as follows:1、We propose a Chinese named entity recognition model based on syntactic information enhancement: Kcr-FLAT.We use the Flat-Lattice Transformer(FLAT)model as the base model to take advantage of its lightweight and parallel computing.In terms of syntactic enhancement,three types of syntactic information are extracted by a key-value memory network and fused through the attention mechanism.Then,the syntax information and vocabulary information are fully fused through Cross-Transformer.Finally,the naming regularities for the same type of entities are extracted through an inner regularity perception module,improving the accuracy of named entity recognition.Using the F1 score as the evaluation index,the experimental results of the proposed model on MSRA,Resume,and Weibo datasets is 96.51%,96.81%,and 70.12%,respectively,basically achieving the expected performance.2、We propose a multi semantic fusion model: MSFNet.This model is based on the general BiLSTM-CRF network structure,establishing a low coupling adaptive embedding layer to enhance the input by introducing various semantic knowledge such as Chinese character radical as glyph information and mutiple word segmentation,thereby improving the accuracy of Chinese named entity recognition.In terms of glyph enhancement,a glyph graph is constructed between radicals,characters,and words,and a graph attention network is used to integrate information from other adjacent character nodes and the radical nodes contained in the character,enabling the model to learn the category information of Chinese characters.In terms of vocabulary enhancement,we use self-attention mechanism to fuse multiple segmentation results to better integrate boundary information of vocabulary and reduce errors caused by a single tokenizer.Finally,the vocabulary information and glyph information are dynamically weighted and combined through a gate mechanism,enabling the model to better utilize the lexicon information and glyph information.Using F1 score as the evaluation indicator,the experimental results show that the model has promising recognition performance in specific scenarios of Chinese named entity recognition datasets such as MSRA and Resume,reaching 96.93% and 97.56% respectively.