Research On Chinese Named Entity Recognition Method Based On Multi-Granularity Feature Fusion

In recent years,the rapid development of Artificial Intelligence(AI)and Natural Language Processing(NLP)technology has significantly transformed human lifestyles.The emergence of ChatGPT and Wenxin Yiyan are leading the way in natural language processing research,particularly in the areas of language comprehension and question-and-answer generation systems,which heavily rely on natural language processing and information extraction technologies.Named Entity Recognition(NER)stands out as one of the most crucial sub-technologies in these technologies.The quality of NER results has a profound impact on downstream tasks such as question-and-answer systems,machine translation,and knowledge base construction.Consequently,NER has garnered significant attention from both academia and industry due to its research significance and practical value.With the advancements in artificial intelligence and deep learning technologies,named entity recognition has achieved a considerable level of proficiency.However,Chinese named entity recognition,which started relatively later,still faces several challenges.Particularly,there are issues in effectively utilizing multi-granularity features such as words,characters,radicals,and cross-sentence context within the sequence.Problems include the underutilization and exploration of word information and radical-level information in Chinese text,as well as the difficulty in effectively leveraging cross-sentence contextual information within the sequence.To address these challenges,this study investigates the extraction and fusion of multi-granularity information,including words,characters,radicals,and cross-sentence context,within the sequence.The primary research focus and innovations of this paper are as follows:(1)An Exp-Soft Lexicon lattice model incorporating radical-level information is proposed in this thesis.A radical-level information extraction module is employed to extract and exploit pictorial information from sequential characters.This involves extracting radical information,construction information,and writing position sequences of 4719 commonly used Chinese characters.Subsequently,these sequences are transformed into low-dimensional dense vectors using word2 vec representation.Text Convolution and pooling are applied for feature extraction.Additionally,the original Soft Lexicon method is extended to provide a more detailed division of position information for intermediate groups,reducing the loss of relative position information of word characters in the Soft Lexicon method.This alleviates the problem of the model’s performance declining significantly as the entity length increases.(2)This thesis proposes a NER method that incorporates sliding cross-sentence contextual information.The sentence-level information extraction module first utilizes the Star-Transformer model to extract global information representation of sentences in a sequence based on BERT encoding.This representation is then stored in a vector lookup representation is dynamically incorporated into the current sequence of sentences during computation.This overcomes the limitation of previous methods,which could only incorporate fixed-length characters.Furthermore,by utilizing sentence global information representation,the cost of introducing cross-sentence context information is reduced from the sentence level to the character level.Compared to methods that do not incorporate cross-sentence context,this approach achieves training efficiency with minimal degradation while exhibiting superior performance as the length of the introduced cross-sentence context increases.This study conducts a comparative analysis of the two proposed models,along with a baseline model,on publicly available datasets from various domains.Precision,recall,and F1 score are employed as evaluation metrics for model performance.The experimental results provide evidence of the effectiveness of the proposed models that integrate multi-granularity features.