Named entity recognition for power data based on lexical enhancement and global pointer
Cong Li, Ran Cui, Zeng Dou, Chengbin Huang, Liang Zhao, Yan Zhang, Cong Chen, Congzhe Su, Jia Li, Chang Qu · 2024
Named Entity Recognition for power data refers to the identification of key designative contents, such as equipment names, operation data, etc., from the text of power domain data to achieve the extraction and classification of key information from the perspective of power expertise. Deep neural networks have shown great effectiveness in power system data NER. However, the Chinese power domain data NER suffers from problems such as insufficient training data and a wide variety of data entities. To solve these problems, we propose a power data entity recognition model based on Lexical enhancement and Global pointer. Firstly, the model uses the Lexical enhancement method to merge the lexical information into the vector representation of each character, then uses the RoBERTa pre-trained model to receive the vectors from the input representation layer and further extract the features, and finally uses the Global pointer method for the entity recognition. In addition, we have experimented on a self-constructed Chinese power named entity recognition dataset, the result of the experiment indicated that the F1 values were higher than several other named entity recognition methods, such as Lattice- LSTM, SoftLexicon+BiLSTM+CRF, CAN and RoBERTa+GP, with improvements of 2.54%, 0.13%, 3.80%, and 0.42%, respectively.