The Research Of Text Sentiment Analysis Based On Deep Learning

Text sentiment analysis is one of the most active research topics in natural language processing,which focuses on the sentiment polarity of given text by analyzing,processing,and summarizing contextual information.In recent years,deep learning has achieved great success in the field of natural language processing.The sentiment classifiers based on deep neural networks can overcome the shortcomings of tradictional methods by means of alleviating the workload of feature engineering.In current researches,however,there are still a series of problems in deep learning sentiment analysis models,such as it is difficult to make full use of the sentiment resources in the given texts,the training time of model is long due to the expensive computation,it is unable to fully extract the long dependency relationship between different sentences in long text.To adjust the above problems,we propose four novel deep learning sentiment classification methods,which improve the algorithm or architecture of model based on deep neural networks.The main research is outlined as follows:i.We propose a novel sentiment classification method for short text sentiment analysis based on multi-channels convolutional neural networks(MCCNN).The model can promote the full use of sentiment features in the short text via different types of feature embedding in the input layer of convolutional neural networks,and the model can indicate the degree of importance of diffrent words in the sentence by means of the position feature.Afterwards,a multi-channels architecture based on convolutional neural networks will be applied to extract sentiment information of the given short text from different patterns.ii.In order to make full use of the sentiment information of specific target and reduce the time cost of the training process,we propose a multi-attention convolutional neural networks(MATT-CNN)for target-based sentiment classification.This approach can capture deeper level sentiment information and distinguish sentiment polarity of different targets explicitly through a multi-attention mechanism without using any external parsing results.Meanwhile,this approach is evidently faster than current attention-based method because of the parallelization input of the model.iii.We propose a gated hierarchical neural model of combining regional long short-term memory and gated convolutional neural network(RLSTM-GCNN)for the task of target-based sentiment analysis.The approach can focus on the local information of the specific target in a sentence by means of regional long short-term memory.At the same time,the RLSTM-GCNN uses the gated convolutional neural networks to extract sentiment features of specific target across the whole sentence,and controls the transmission of sentiment information through a gated operation,so that the model is able to effectively infer the sentiment polarity of different targets in the same sentence.iv.In view of the problem that attention-based neural models ignore the long-distance dependency of aspect across the entire long text and cannot take full advantage of context relations of different sentences in the same review,we propose a hierarchical model of combining regional convolutional neural network and hierarchical long short-term memory(RCNN-HLSTM)for the task of aspect-based sentiment classification on long text customer review.This approach can extract in-depth sentiment information of aspect in the individual sentence and across the whole review via hierarchical long short-term memory,and capture the long dependency of different sentences in the same long text review by means of regional convolutional neural networks.Such hierarchical is able to consider both intra-sentence and inter-sentence relations of specific aspect in the long text review.