Incorporating token-level dictionary feature into neural model for named entity recognition

For named entity recognition, incorporating dictionary feature into the deep neural model can obtain better performance. However, the behavior of deep models which incorporate dictionary features is not clear because analyzing this behavior is difficult. Besides, using the dictionary features which depend on a dictionary may train a poor model. In this work, we present an incorporating token-level dictionary feature method which use a labeled dataset rather than an external dictionary. This approach makes it easier to quantify the effect of dictionary features and decouples dictionary features from the external dictionary during the training stage. Additionally, we conduct several experiments which consist of three parts. First, we show the effectiveness of our proposed method. Second, we study the behavior of incorporating dictionary features from two perspectives: during the training and the testing stage. Third, we research the embedding of dictionary features. (C) 2019 Elsevier B.V. All rights reserved.