Named Entity Recognition Method Based on ERNIE2.0-BiLSTM-AT-CRF-FL
Qing Li, YuKe Lv, Yanping Zhou · 2023
To address the problem that the named entity recognition task cannot address the phenomenon of multiple meanings of words in context and unbalanced label classification when dealing with Chinese words, the ERNIE2.0-BiLSTM-AT-CRF-FL entity recognition method is proposed, and the pre-training model adopts enhanced language representation with informative entities 2.0 (ERNIE2.0). The model can learn dynamically by combining the context of words to obtain dynamic semantic features of words and solve the phenomenon of multiple meanings of words; Then the word vector is input to the bi-directional long short-term memory (BiLSTM) neural network to extract features, and the weight size of each word is calculated by the soft attention mechanism (AT) layer; Finally, the entity labels are obtained by decoding in the conditional random field (CRF), and by using the focal loss function (FL) to alleviate the label classification imbalance. The proposed algorithm model is applied to the MSRA corpus for validation, and the experimental results show that the model achieves an F1 value of 95.39% which can better handle the Chinese-named entity recognition task.