双向循环网络中文分词模型

Bidirectional Recurrent Networks for Chinese Word Segmentation

针对统计方法的中文分词模型主要依赖于特征工程,难以捕捉句子中长距离依赖关系等问题,提出一种双向循环网络中文分词模型.为能有效获取待分类字符的上下文特征,避免局部窗口大小的限制,使用长短时记忆网络(Long Short-Term M emory Neural Netw ork,LSTM)作为神经网络隐藏层,同时增加一层反向LSTM抽取字符的将来信息特征.提出一种语言模型预训练的网络权值初始化方法,该模型同时得到中文字符embeddings分布式向量特征.在标准分词数据集上测试表明该模型取得比以往统计标注方法更好的效果.通过对比实验结果发现深层神经网络能提取出不逊于人工总结的分词特征.

Focusing on the issue that the statistical methods for Chinese word segmentation （ CWS ） is mainly dependent on hand-craft feature engineering,and difficult to capture long range dependences in sentence. This paper proposes a novel method for CWS based on Bidirectional Recurrent Neural Networks （ BRNN ）. In order to obtain context feature effectively and avoid limitation of local window size, Long Short-Term Memory （ LSTM ） is used as the hidden layer of the architecture, as well as future information captured by reverse direction LSTM. In this model, we propose a new approach for the initialization of network weights and Chinese character embeddings,which is based on language model method. The experiments demonstrate this approach improves the accuracy of CWS in standard dataset. By comparing the experimental results, it is found that the deep neural network can automatically extract features for CWS, which not inferior to traditional feature engineering.