Performance Comparisons of Bi-LSTM and Bi-GRU Networks in Chinese Word Segmentation

Abstract

The Bi-directional Long Short-Time Memory (Bi-LSTM) neural networks can effectively use contextual information in both directions when comparing with the LSTM neural networks. It is more advantageous to extract text information in the word segmentation process. Based on the Bi-LSTM neural networks, Bi-directional Gate Recurrent Unit (Bi-GRU) neural networks can reduce the amount of calculation and cost less time. In this paper, we use Bi-LSTM to segment Chinese words and find the best hyper-parameters. BI-GRU and Bi-LSTM are employed to conduct a comparison experiment of Chinese word segmentation performance. Experimental results prove that the Bi-GRU neural networks have a faster speed, but a simpler structure than the Bi-LSTM neural networks. Meanwhile, it has no loss of word segmentation accuracy.