Recognition of News Named Entity Based on Multi-Level Semantic Features Fusion and Keyword Dictionary
<p>Yi Ren, Yu Liu</p> · Academic Journal of Computing & Information Science · 2024
Most of the current methods used existing models for named entity recognition tasks, but this could only obtain character vectors and could not solve the problem of polysemy. This study proposed a new model based on multi-level semantic features fusion and dictionary of keyword to solve this problem. This method first uses a keyword dictionary for random entity replacement to achieve data augmentation; then, it utilizes the pre-trained BERT model to transfer prior knowledge to this task to obtain multi-level semantic features; Secondly, in order to obtain more comprehensive sequence information, the vector is input into the multi-semantic feature fusion layer to extract global information; Finally, after correcting the results with the CRF, the output is obtained. Compared with traditional models such as BiLSTM-CRF and BERT-CRF, this model has achieved good results on news domain datasets, with an accuracy rate of 94.95% and an F1 value of 94.99%.