Named Entity Recognition Based on Reinforcement Learning and Adversarial Training

Abstract

In this paper, we propose a new model that combines reinforcement learning and adversarial training to exploit the data generated by distant supervision for named entity recognition. Our model can not only reduce the influence of noise in generated data, but also find more informative instances for training. In the pre-training stage of the model, in order to make full use of the data generated by distant supervision, we use reinforcement learning to select reliable instances to pre-train a classifier. In the training stage of the model, we introduce the adversarial training mechanism, which can not only find more reliable instances to enhance the ability of the classifier, but also use noise data to improve the ability of the model to resist noise. To evaluate the performance of the model, we conduct experiments on two public datasets, Species800 dataset in biology and EC dataset in e-commerce domain. The experimental results show that in Species800 dataset, the F1 score of our model is 1.68% higher than that of baseline, and in EC dataset, the F1 score of our model is 6.32% higher than that of baseline. Compared to the state of art models, our model can achieve comparable performance just using word2vec embedding.