Fine-grained Sentiment Analysis Of Online Comments Based On Deep Learning

With the rapid development of the Internet and e-commerce platforms,online transactions have become a more and more way for netizens to consume.At the same time,consumers’online reviews of products have also increased dramatically.These online reviews can provide consumers with decision-making opinions,and can also provide businesses with improved directions,which are important large-scale text data.Aiming at the above application background,the thesis proposes a fine-grained sentiment analysis model based on deep learning,which can simultaneously predict the emotional tendencies of many aspects in online reviews.The main work of the thesis is as follows:Firstly,the text set is preprocessed,and the BiLSTM-CNN-CRF model is proposed to implement Chinese text segmentation.The text is then statistically analyzed and solutions are proposed for the problem of unbalanced data sets.Then the data enhancement of Chinese text is realized.The experimental results show that data enhancement can improve the F1 score of fine-grained emotional prediction results.Secondly,a method for combining online representation of word embedding and character embedding is proposed.The use of a single word embedding indicates that there are certain drawbacks:the quality of the word segmentation directly affects the effect of the word embedding model and the problem that the model will appear unregistered words.In this case,the word embedding cannot effectively characterize the words.The thesis realizes the joint representation of word embedding and character embedding,which has the advantages of two kinds of embedding representation,reduces the influence of word segmentation error and avoids OOV problem to some extent,and can fully learn the semantic representation of words.Experimental results show that the effect of embedding joint representation is better than single embedding representation.Thirdly the improved Skip-gram LDA model is implemented,which can capture the more fine-grained word co-occurrence relationship and make up for the defect that the LDA topic model ignores the order between words and words.In the thesis,the improved Skip-gram LDA model is used to obtain the topic of online commenting.The topic obtained is the evaluation aspect in the fine-grained sentiment analysis task.Then,based on the gradient boosting decision tree algorithm,the extraction of aspect keywords is realized.Then the aspect keyword is used as the input of the fine-grained sentiment analysis model.The experimental results show that this can make the model better predict the emotional tendency corresponding to the evaluation.Finally,the overall model of the fine-grained sentiment analysis task is implemented.The online comment and aspect keywords are used as the input of the model,and the LSTM-CNN based on the attention mechanism is used to extract the text features,and the tanh-relu gate controls the flow of emotional information in each aspect,through the pooling layer and the fully connected layer.At the same time,the multi-faceted emotional tendency was predicted.Finally,the fine-grained sentiment analysis model designed in the thesis achieved a F1 score of 0.715.