IDCNN-CRF-based domain named entity recognition method
Bihui Yu, Jingxuan Wei
Abstract
Entity recognition is the foundation of natural language processing. For traditional methods, a large number of manual annotations are required, and the accuracy of the model is low and the speed is slow. The word vector model cannot recognize unregistered words well. This paper proposes an entity recognition method based on character embedding, Iterated Dilated Convolutional Neural Networks (IDCNN) and Conditional Random Fields (CRF). Combining the characteristics of iterative dilation convolutional neural network GPU parallel operation and long-term and short-term memory, the ability of word vectors to express the meaning of unregistered words, and the excellent learning ability of conditional random fields for labeling rules, a character+IDCNN+CRF named entity is constructed Identify the model. Based on the corpus in the field of military equipment, the experiment shows that the method can distinguish the equipment name and organization name excellently in a certain dimension character vector. The F-1 value in the test corpus exceeds 94%. For military equipment domain entity recognition has a better effect, and the prediction speed has been greatly improved compared to before.