Research And Application Of Fine-grained Text Sentiment Analysis For Product Reviews

With the increasing popularity of online shopping,people tend to publish some subjective comments on goods or services while choosing online shopping.These comment statements not only reflect the views and attitudes of consumers,but also give certain feedback to the providers of goods or services.Deeply digging the information value in the review text can not only provide positive feedback for businesses,but also provide decision-making reference for consumers,which has important research significance.The text sentiment analysis algorithm in natural language processing can perform positive,negative,and neutral sentiment analysis on product reviews.The overall text sentiment analysis of the sentence ignores the emotional polarity of the specific attributes of the product,and the analysis granularity is relatively coarse,so it needs to be specific to the product attributes for fine-grained sentiment analysis.The fine-grained text sentiment analysis of product review sentences can be divided into text sentiment analysis of specific goals and text sentiment analysis of aspect categories.The main work of this thesis is as follows:(1)In the text sentiment analysis of a specific target,when the target word is composed of multiple words,the representation method of taking the average vector of multiple words is easy to cause semantic confusion.This thesis designs an internal attention mechanism based on the target word to better represent the target words;At the same time,a bilateral interactive attention mechanism is designed based on the target word and the context word of the text,which strengthens the semantic interaction between the target word and the context word of the text,and better infers the emotional label of the target word;the neutral emotion in the comment sentence have emotional ambiguity,which is an important factor affecting the accuracy of the model.This thesis designs a label smoothing algorithm to process neutral comment sentences.The experimental results on commonly used datasets show that the algorithm proposed in this thesis has achieved better results than various models combining machine learning and attention mechanisms.(2)In the text sentiment analysis algorithm of the aspect category,the semantic abstraction of the comment sentence is aimed at,and the semantic connection between the aspect category words and the whole sentence of the comment sentence is not strong,and the semantic information characteristics of the comment sentence cannot be extracted effectively,and the classification accuracy is limited.This thesis first constructed auxiliary sentences based on the aspect category words as the input of the Bert model.Based on the Bert model obtaining the semantic information of the comment sentences and auxiliary sentences,the design module was used to fuse the semantic vector and output result of the middle layer "CLS" of the Bert model.The semantic vector of the middle category words further strengthens the connection between the review text and the category words,and extracts deeper semantic information to send to the classifier.The experiment has achieved better classification results than the basic neutral network combined with attention-based models and the pure Bert model,which verifies the effectiveness of the feature fusion method proposed in this thesis.(3)Based on the above research,a sentiment analysis system for product reviews is constructed here.The pre-training model is introduced through the processing of the product review corpus,and then the review text is pre-processed,and the specific target text sentiment analysis algorithm and aspect category text sentiment analysis algorithm are pre-trained,the algorithm model is called,and then WEB-side text sentiment analysis is finally realized.