An Attention-Enhanced LSTM Model for Chinese Educational Named Entity Recognition

This paper proposes an attention enhancement and parameter tuning-based Named Entity Recognition (NER) method under resource constrained scenarios in the Chinese education domain: BERT-BiLSTM-Transformer-CRF. This is the first model of its kind proposed in the field of NER, and it solves the problem of gradient vanishing and exploding in the baseline model of BERT-BiLSTM-CRF in long sequences processing. The model significantly enhances the ability of capturing long range contextual dependencies by the introduction of the Transformer. At the same time, our work represents a pioneering effort in the design of models for the Chinese educational NER tasks. Experiments on EduNER dataset show that the proposed model improves F1 score by 0.65 to 4.73 in comparative experiment and in ablation experiment improve Precision by F1 score by 0.27 to 6.92. The results highlight the efficacy and superiority of incorporating Transformer mechanisms in advancing Chinese educational NER under resource-constrained scenarios.