Aspect-level Emotion Analysis Of Online Comments Based On WMAB And CNN

In recent years,the amount of online commentary data has grown exponentially.In order to obtain valuable information from these review data quickly,scholars began to mine views on these data,opinion mining is the emotional analysis of sentences with emotional tendency.According to different levels of comment text mining,opinion mining techniques are divided into three levels,namely text-level,sentence-level and aspect-level.Mining aspect-level opinion of online commentary data can help merchants and consumers to obtain detailed information on all aspects of the evaluation object.For the main problems existing in perspective aspect-level mining technology,this paper analyzes and improves the technology from two aspects,the aspect-level view information mining and the aspect-level sentiment classification.Firstly,the traditional unsupervised learning method is mainly based on the principle of word frequency statistics,so it is easy to lose hidden words and reduce the effect of opinion information extraction.This paper proposes a new algorithm WMAB(Word2vec based MAB).Based on the idea of MAB(Multi-Aspect Bootstrapping)algorithm,this method integrates the semantic information between Word2vec word vectors,combing the semantic similarity between word vectors and bootstrapping,and then selects more accurate viewpoint words by calculating the aspect importance score of candidate viewpoint words.WMAB algorithm can effectively overcome the shortcomings of traditional methods,which are mainly based on the idea of word frequency statistics,so as to improve the accuracy of opinion word mining in the aspect of network comments.Secondly,for the text feature matrix spliced by simple word vectors is insufficient aspect-level feature extraction and unclear key information,this paper proposes a text feature matrix construction method that combines aspect-level weights.Aspect-level weight includes opinion word importance scorescore_i(w)and attribute class weight.The method highlights the viewpoint word vectors corresponding to each evaluation aspect in the comments according to different aspect weights,enhances the aspect-level differences of the comments text,and improves the accuracy of supervised learning algorithm for aspect-level emotional classification.The paper uses the multi-granular convolution neural network Multi-CNN to verify the above viewpoints,and proposes a pooling method for merging the maximum value and residual information for the problem of serious data loss of the maximum pooling function of the Multi-CNN pooling layer.So the effect of Multi-CNN on the aspect-level emotional classification of online comment data can be further improved.Finally,two sets of experiments were performed separately.The first set of experiments is used to verify the validity of the WMAB algorithm which fuses semantic information between Word2vec word vectors for aspect-level view information mining.The second set of experiments is used to verify that the text feature matrix of the fusion aspect importance score and attribute class weights improves the effect of Multi-CNN on aspect-level sentiment classification,and proves the effectiveness of pool layer-improved Multi-CNN.