摘要
传统的中文分词方法是一种基于单词标注的传统机器学习方法,但学习方法需要人工配置和提取中文文本的特征。缺点是同义词库维度较高且CPU训练模型较长。本文针对以上问题进行了研究,构建了内嵌条件随机场的长短时神经网络模型,使用长短时神经网络隐含层的上下文向量作为输出层标注的特征,使用内嵌的条件随机场模型表示标注之间的约束关系采用双向LSTM和CRF相结合的训练方法进行特定领域知识点的中文分词。对中文分词测试常用语料库的实验比较表明,基于BLSTM和CRF网络模型的方法可以获得比传统机器学习方法更好的性能;使用六字标记并添加预训练的字嵌入向量可以实现相对较好的分词性能;BLSTM-CRF网络模型方法更易于推广并应用于其他自然语言处理中的序列标注任务。
The traditional Chinese word segmentation method is a traditional machine learning method based on word annotation, but the learning method requires manual configuration and extraction of Chinese text features. The disadvantage is that the thesaurus has a higher dimension and the CPU training model is longer. In this paper, the above problems are studied, and the long-term and short-term neural network model of embedded conditional random field is constructed. The context vector of the hidden layer of the long-term neural network is used as the feature of the output layer annotation, and the embedded conditional random field model is used to represent the annotation. The constraint relationship between the two-way LSTM and CRF combined training methods for Chinese word segmentation of specific domain knowledge points. The experimental comparison of the common corpus of Chinese word segmentation test shows that the method based on BLSTM-CRF network model can obtain better performance than the traditional machine learning method;using six-character mark and adding pre-trained word embedding vector can achieve relatively good segmentation Performance;BLSTM-CRF network model methods are easier to generalize and apply to sequence annotation tasks in other natural language processing.
作者
周海华
曹春萍
ZHOU Hai-hua;CAO Chun-ping(University of Shanghai for Science and Technology, School of Optical-Electrical and Computer Engineering, shanghai 200082, China)
出处
《软件》
2019年第2期1-5,共5页
Software
基金
国家自然科学基金项目(61402288)
上海市自然科学基金(15ZR1429100)