Research On Chinese Word Segmentation Based On Neural Network

In recent years,with the continuous development of the Chinese Internet world and the deepening of artificial intelligence research,Chinese natural language processing technology has become increasingly important.Word segmentation is a key technology of Chinese natural language processing and is indispensable in many applications.It is an effective way to treat Chinese word segmentation as a kind of character-based sequence labeling problem so as to adopt machine learning to deal with it,which is called Character-based Chinese word segmentation.The character-based Chinese word segmentation treats given Chinese statements as sequences of Chinese characters,and then assigns an label to each character element in the sequences.However,the traditional statistical machine learning method requires a large number of features to be carefully designed by human.Further improvement of the model effect is constrained.However,on the one hand,the inherent sequence nature make the training time of LSTM longer,which is not conducive to the practical application of neural network model.However,on the one hand,the inherent sequence nature make the training time of memory network longer,which is not conducive to the practical application of neural network model.On the other hand,when dealing with Chinese word segmentation using neural network methods including short-short time memory network,many works still need necessary feature engineering to enhance the ability of the model to capture local features.In view of these,this paper mainly does the following work:First,in view of the difficulty in training the existing LSTM-based models,this paper no longer use the short-short time memory network to extract context features,but introduces the self-attention mechanism used in the sequence-to-sequence tasks into the Chinese word segmentation task as a new feature extractor.To our best knowledge,this is the first time to introduce self-attention mechanism into Chinese word segmentation task.Secondly,in order to verify the importance of local features in Chinese word segmentation task,this paper designs a local self-attention mechanism.Local self-attention mechanism changes the original method of self-attention mechanism to model global information and only pays attention to local context information in a certain range.Compared with the original self-attention mechanism,the new local self-attention mechanism does not need additional parameters.Thirdly,in order to enhance the ability of self-attention mechanism to capture local information,this paper combines Convolutional Neural Network(CNN)with self-attention mechanism.The aim is to use CNN to supplement the excellent ability of extracting local features for self-attention mechanism.This method takes the encoder block as the basic component.In each encoder block,the convolution network is used to extract local features,and then the extracted local features are input into the self-attention mechanism to capture the global information of the input sequence.This paper conducts experiments on two benchmark data sets.Experimental results show that without any features engineering and extra dictionary,the proposed local attention mechanism fits faster than the original self-attention mechanism;Compared with the Chinese word segmentation model based on long-term and short-term memory network,the model based on combining convolution neural network with self-attention mechanism has certain advantages in computational efficiency and segmentation effect,which provides new ideas and methods for Chinese word segmentation task.