Deep Semi-Supervised Learning And Its Application In Sentiment Analysis

With the development of information technology and network technology,the amount of data explosive growth.And in this huge amount of data,valuable information hidden in it.Data mining is the process of figuring out how to get these information.People use one of the important tool in data mining: machine learning,to process big data,like sift out gold from sand.As an important branch of machine learning: deep learning has been widely used by researchers in recent years.However,training a deep supervised learning model need a mass of good tag samples,and tagging labels is a time-consuming and laborious process.In the meantime,unlabeled samples are relatively easy to obtain in many tasks.Therefore,how to effectively use unlabeled data to improve generalization performance has become a key problem in the field of machine learning.Semi-supervised learning can be in the case of little tagging sample,reduce dependence on large scale marked sample by joining in a large number of unlabeled samples in the process of training.In recent years,the consistency hypothesis in deep semi-supervised learning has become one of the hotspot.The so-called consistency hypothesis applies data augmentation to the training process,then it leveraging the idea that a classifier should output the same class distribution for a training sample even after it has been augmented,thus improved model's robustness.However,it is found that in the process of calculating the consistency loss,the method which base on the consistency hypothesis only considers adding perturbation to each sample,but ignores the connection between them,and may further losing the of structured information of the data set.Therefore,we proposes a deep semi-supervised learning algorithm combining consistency regularization and manifold regularization,and apply it to the task of sentiment analysis.The specific work content is as follows:(1)Aiming at the shortcoming of losing data set structure information based on consistency semi-supervised method,we propose a deep semi-supervised learning algorithm combining consistency regularization and manifold regularization.The algorithm not only imposes consistency constraints on the model,but also construct a graph by the data set and add smoothness loss to the model.On the one hand,this process will make the model in the neighbor of each sample is smooth.On the other hand,the smoothness between the points are also guaranteed.The experiment results show that our algorithm improved the model accuracy compared with other semi-supervised algorithms.(2)Faced with the complexity of the acquisition of labeled samples and the easy acquisition of unlabeled samples in sentiment analysis,we apply the semi-supervised algorithms paper proposed and a variety of classic semi-supervised learning algorithm to the sentiment classification of sentiment analysis,constructing a B/S structure of semi-supervised sentiment analysis system.The main functions of the system include: text preprocessing,feature extraction and algorithm implementation.At the same time,the system introduce the application of semi-supervised learning in sentiment analysis.