基于BI-LSTM-CRF模型的中文分词法

Chinese Word Segmentation Based on Bi-directional LSTM-CRF Model

递归神经网络能够很好地处理序列标记问题,已被广泛应用到自然语言处理(NLP)任务中。提出了一种基于长短期记忆(LSTM)神经网络改进的双向长短期记忆条件随机场(BI-LSTM-CRF)模型,不仅保留了LSTM能够利用上下文信息的特性,同时能够通过CRF层考虑输出标签之间前后的依赖关系。利用该分词模型,通过加入预训练的字嵌入向量,以及使用不同词位标注集在Bakeoff2005数据集上进行的分词实验,结果表明:BI-LSTM-CRF模型比LSTM和双向LSTM模型具有更好的分词性能,同时具有很好地泛化能力;相比四词位,采用六词位标注集的神经网络模型能够取得更好的分词性能。

Recurrent neural network had been broadly applied to natural language processing（NLP） problems,because they deal well with the problem of sequence labeling. In this paper, we propose to use bidirectional LSTM CRF（BI-LSTM-CRF） model for Chinese word segmentation, which is based on long short-term memory（LSTM）units. This model not only can keep the contextual information in both directions,but also through the CRF layer to consider the dependency between the output tag. By using different tag set and adding pre-trained character embeddings, and using the model in the Bakeoff2005 data set on the word segmentation experiment results show that：BI-LSTM-CRF model has better segmentation performance than LSTM and bidirectional LSTM model,and has good generalization ability;Compared with the four-tag-set,the neural network model with the six-tag-set can achieve better segmentation performance.