Named Entity Recognition Model Based on the Fusion of Word Vectors and Category Vectors

Named entity recognition (NER) in deep learning mode heavily relies on the processing and analysis of text vectors. This paper introduces an NER model based on deep learning techniques, emphasizing the fusion of word vectors and category vectors to enhance text processing and analysis capabilities. The model consists of two core components: a Fusion Vectors Generation (FVG) module and an NER module. In the FVG module, word vectors are derived from the BERT pre-training model, encapsulating the semantic information of the text. Additionally, category vectors are introduced to represent the entity type associated with the text. The relationship between word vectors and category vectors is learned using the BERT-BiLSTM architecture, and fusion vectors are then generated by combining these two types of vectors. On the other hand, the NER module utilizes the BiLSTM-CRF structure to extract contextual features from the fusion vectors and optimize the predicted label sequences. Consequently, a mapping between the fusion vectors and classification labels is established, facilitating the NER process. Validation on three NER datasets—Weibo, Youku, and Chinese literature—demonstrates that the proposed model is both feasible and effective. Compared to a baseline model that solely relies on word vectors, the proposed model exhibits enhanced NER performance, achieving improvements in
$F1$
scores by 5.05%, 1.53%, and 1.81% respectively on the Weibo, Youku, and Chinese literature datasets.