Sentiment Analysis Method Research Based On Deep Learning

The traditional methods used to solve the problem of text sentiment analysis include unsupervised methods based on emotional dictionary and artificial judgment rules,and supervised methods based on machine learning.In the case of the amount of data is not large or semantics is not rich enough,These methods can achieve some effects.But with the increasing amount of the data and the expression is more and more rich,the traditional methods has been unable to effectively solve this kind of problem.New method needs to be put forward.Based on the characteristics of text sentiment analysis and the deep learning algorithm for a wide range of applications in the current,this paper focuses on the method for sentiment analysis based on depth learning.Convolution neural network and recuesive neural network are the two mainstream models in deep learning.The former can extract local features from the data,and the latter can analyze the time-series data effectively,and have strong context generalization ability.The work of this paper includes the following two aspects:1.Sentiment analysis model based on convolutional neural networks and word adjacent features.At present,the methods based on convolutional neural network has achieved good performance in the task of sentiment analysis.This method mainly uses word vectors as the input of the network,but in the convolutional process,the word vector can only characterize a single word,but does not contain contextual information,this is not conducive to the continuity of information transmission,and convolution may disrupt the sequence of word vectors in the local context.Aiming at this problem,this paper proposes a convolutional neural network model based on word adjacent features in the third chapter,which allows each word vector to carry the characteristics of its adjacent words in the convolution process,so that we can both ensure the continuity of information transmission and the sequence of word vectors in the local scope.The experimental result shows that the prediction accuracy rate on the sentiment analysis task of COAE2014 and COAE2015 reached 89.43%and 85.61%respectively,which indicates that the proposal model is actually feasible and effective.2.Sentiment analysis model based on recursive neural networks and artificial judgment rules.The conventional rule-based approach is from a linguistic perspective,but these method involves formulating sentiment dictionary and amount of judgment rules.However,the method based on recursive neural network can learn the prior knowledge from a large number of unlabeled corpus.After referring to these two methods,we present a novel sentiment analysis model based on recursive neural network and artificial judgment rules in the fourth chapter.Firstly,the sentiment polarity of multiple clauses that make up the original sentence are calculated by the parallel recursive networks,then the outputed polarity of the clauses are merged by polarity fusing rules to calculate the sentiment polarity of the original sentence.The advantage of this model on the one hand is that the recursive neural network model incorporates artificial judgment rules,so that artificial experience is effectively used.On the other hand,the recursive neural network replaces the sentiment dictionary,and this can avoid the limitations of the emotional lexicon.The classification accuracy on the SST-C2 and SST-C3 datasets are 87.8%and 81.6%respectively,and the overall performance is better than the mainstream classification model,which indicates that the proposed model is not only novel but also actually feasible and effective.