Research And Application Of Text Sentiment Analysis Based On Deep Learning

Sentiment analysis is one of the classic research branch in the field of Natural Language Processing.With the maturity of the Internet and e-commerce,people have accustomed to shopping and ordering meals on various online platforms.After consumption on the online platform,users will make comments on the commodity.These comments information is growing day by day and has extremely high research value.By analyzing and mining these information,we can grasp users' preferences and consumption needs,and provide reference for other consumers' consumption behavior.Businessmen can also improve their products quality according to consumers' consumption needs.Traditional sentiment analysis methods mainly include rule-based and machine learning methods.The rule-based approach requires the construction of the sentiment lexicon.The effect of sentiment polarity classification depends on the quality of sentiment lexicon,and it is very difficult to construct a universal and interdisciplinary sentiment lexicon.However,machine learning methods need to extract features,which usually can't represent the semantic information.Therefore,this thesis mainly studies the application of deep learning method in sentiment polarity classification.The work of this thesis is as follows:(1)For the task of coarse-grained sentiment analysis at the sentence level,this thesis proposes a model to enhance sentence representation ability from multiple perspectives(global maximum pooling,global average pooling and attention mechanism).Only using maximum pooling can extract the important information in the sentence,but at the same time,some valuable information will be lost.Therefore,this thesis combines global max pooling,global average pooling and attention mechanism to enhance sentence representation ability.This model has a good Accuracy value of 82.41% and 86.59% on the Customer Review dataset.At the same time the model has a good F1 score value of 0.822 and 0.862 on the Customer Review dataset,achieving satisfactory results and outperforming other baseline models on both datasets.(2)This thesis proposes a model based on self-attention mechanism for aspect-level fine-grained sentiment analysis task.In this model,two modules based on self-attention mechanism are used for sentence representation,and the information of 20 specific aspects is gradually obtained,and the sentiment tendency labels of 20 aspects are finally outputted at one time.This model has achieved good results in the 2018 AI Challenger fine-grained sentiment analysis dataset,with an average F1 value of 0.7084 and an average Accuracy value of 88.65%,which is better than other baseline models.(3)Applying the proposed model to the sentiment analysis of online catering reviews,it mainly includes three functions of data capture,coarse-grained sentiment analysis and aspect-level fine-grained sentiment analysis,which verifies the validity and practicability of the model proposed in this thesis.