基于门循环单元神经网络的中文分词法

A Gated Recurrent Unit Neural Network for Chinese Word Segmentation

目前,学术界主流的中文分词法是基于字符序列标注的传统机器学习方法,该方法存在需要人工定义特征、特征稀疏等问题.随着深度学习的研究和应用的兴起,研究者提出了将长短时记忆(long short-term memory,LSTM)神经网络应用于中文分词任务的方法,该方法可以自动学习特征,并有效建模长距离依赖信息,但是该模型较为复杂,存在模型训练和预测时间长的缺陷.针对该问题,提出了基于门循环单元(gated recurrent unit,GRU)神经网络的中文分词法,该方法继承了LSTM模型可自动学习特征、能有效建立长距离依赖信息的优点,具有与基于LSTM神经网络中文分词法相当的性能,并在速度上有显著提升.

Currently, the common method for Chinese word segmentation is traditional machine learning on character-based sequence labeling.However,this method faces disadvantages such as manual feature engineering and sparse features.With the increasing re- search and application of deep learning,researchers have proposed a method by applying long short-term memory （LSTM） to Chi- nese word segmentation task. This method is capable of learning features automatically and capturing long-distance dependence as well.However, this method is complicated,and has defects in speed.Therefore,we propose a gated recurrent unit （GRU） neural net- work for Chinese word segmentation,which are also associated with advantages of learning features automatically and the ability of capturing long-distance dependence.Finally,our method performs comparably well as the LSTM neural network for Chinese word segmentation,and exhibits a great improvement in training and predicting speeds.