Adapting deep belief nets to Chinese entity detection

Abstract

This paper adapts deep belief networks (DBN) to detect entity mentions in Chinese documents. Our results exhibit how the depth of architecture and quantity of unit in hidden layer influence the performance. Different feature combinations are used to show their advantages and disadvantages in DBN for this task. Moreover, we combined Chinese word segmentation systems to alleviate word segmentation error. Token labels are produced independently by DBN which does not concerned what are the token labels before current word. Viterbi algorithm is a good solution to find the most likely probability label path to make DBN be more effective for entity detection. Furthermore, this paper demonstrates DBN is a proper model for our tasks and its results are better than Support Vector Machine (SVM), Artificial Neural Network (ANN) and Conditional Random Field (CRF).