Entity slot recognition based on data enhancement and character-word fusion features

Aiming at the problem that the traditional entity slot recognition model based on character level representation could not make good use of word information，and the information technology (IT) operation and maintenance field would lack enough open data sets，an entity slot recognition method based on BERT_Word2vec_BiLSTM_CRF model was proposed，and the training data set of the model was extended by data enhancement．In this model，character vector representation obtained by bidirectional encoder representations from transformer (BERT) model and word vector representation obtained by Word2vec were fused，context encoding was carried out by bi-directional long short-term memory (BiLSTM) and decoding was carried out by conditional random field (CRF)，and the final sequence annotation result was obtained．Experiment results on the data set provided by an enterprise show that the fusion of word-level features can further improve the recognition performance on the basis of BERT pre-training model，and the F1 value reaches 92.33%. © 2022 Huazhong University of Science and Technology. All rights reserved.