基于最简门单元的循环神经网络分词

为解决长短期记忆(long short-term memory,LSTM)单元循环神经网络结构复杂,训练时间长,标注推理速度慢的问题,结合现有文献分析循环神经网络及其单元结构的理论基础,提出一种基于最简门单元(minimalist gated unit,MGU)的循环神经网络进行中文分词研究。使用MGU单元替换LSTM单元自动提取特征,建立长期依赖信息。在中文分词评测常用语料Bakeoff 2005数据集上进行实验,实验结果表明,MGU网络与LSTM网络精度相当,训练时间减少一半,标注推理速度可提升至3倍。

MGU RNN for Chinese word segmentation

To solve the problem of long training time and low tagging speed of long short-term memory (LSTM) recurrent neural networks (RNN) caused by its complex structure,combined with the current literature,the basic theory of recurrent neural networks and its unit structure were analyzed.RNN based on minimalist gated unit (MGU) was proposed to study Chinese word segmentation.LSTM units were replaced by MGUs to extract features automatically and establish long term dependence.Results of experiments on common used Bakeoff 2005 datasets show that MGU network and LSTM network get similar accuracy,while MGU reduces training time by half,and raises the speed of word tagging to 3 times.