A chinese named entity recognition method for small-scale dataset based on lexicon and unlabeled data

Recently, using lexicon information to improve the performance of Chinese named entity recognition has been proven to be effective. Moreover, the lexicon-based method represented by Lattice-LSTM has also become the mainstream. Although Lattice-LSTM can introduce lexicon information into characters to augment named entity recognition performance, it cannot make good use of unlabeled data, which contains abundant semantic information to assist the network to improve effect. And because Lattice-LSTM introduces much lexicon information, there is currently no suitable way to assign weights to each word. In this work, we propose a method that can effectively introduce lexicon information, which is also simple to implement and can be applied to various networks. Based on the lexicon method, this method uses external unlabeled data to count the word frequency and improved mutual information to represent the weight of the word to introduce lexicon information. And attention mechanism is used to dynamically assign weights to each part of lexicon information. In this method, the fusion of character and lexicon information is processed before the input layer, so that the method has a faster training speed and better versatility. Compared with other methods that are based on lexicon information, this method introduces additional prior knowledge, namely unlabeled data, and achieves better results when the scale of dataset is small. And when combined with the pre-trained language model, the performance is better (the F1 scores on Weibo dataset and Resume dataset are 96.73% and 71.53% respectively). Experimental research shows that our method surpasses many other excellent baseline methods in training speed and performance on two small-scale public Chinese named entity recognition datasets.