Research And Implementation Of User Comment Sentiment Analysis Based On Deep Learning

With the rise of the mobile Internet,people’s food,clothing,housing and transportation are closely linked with the Internet in all aspects.With the rise of social media such as Weibo,We Chat,and forums,more and more netizens are willing to post their knowledge,opinions,and ideas on the Internet.When consumers on the Internet purchase goods and services,they can also easily evaluate the purchased goods or services.These review texts contain people’s consumer opinions and emotional factors.Researching and mining the emotional tendency information in user review texts has extremely high commercial and social value.Traditional sentiment analysis methods mainly include methods based on sentiment dictionaries and methods based on machine learning.In the era of big data,traditional sentiment analysis methods are gradually no longer applicable in the face of massive text information,and methods based on deep learning have gradually become universal choice in sentiment analysis tasks.This thesis mainly focuses on coarse-grained sentence-level sentiment analysis and finegrained aspect-level sentiment.The existing deep learning-based sentiment analysis model is researched and improved,and for the current research trend of sentiment analysis tasks,pretrained language models and the technology of transfer learning is introduced into the sentiment analysis model.The main work of this thesis is as follows:(1)This thesis studies the shortcomings of the current sentence level sentiment analysis.Based on the existing research,this thesis proposes a sentence level sentiment analysis algorithm CBLSAN,which uses multi-channel convolutional neural network and bidirectional Long short-term memory network to obtain the semantic information features of the text.At the same time,it introduces self-attention mechanism,which improves the learning of the key features of the text and improves the accuracy of sentence level sentiment analysis classification results.(2)Research the existing problems of aspect-level sentiment analysis.Based on the existing researches,this thesis proposes an aspect-level sentiment analysis PIMSN algorithm model,which uses the position information of aspect words combined with multi-head self-attention and multiple interactive attention mechanisms to improve the information perception ability of the algorithm model.This model improves the performance of aspect-level sentiment analysis classification results.(3)Aiming at the pre-training-fine-tuning mode that is commonly used in the field of natural language processing,this thesis introduces the BERT pre-training model,and uses the BERT word vector to improve the PIMSN algorithm proposed in this thesis.Using the BERT word vector can obtain more contextual interaction information.This thesis proposes an improved PIMSN-BERT model,which further improves the classification accuracy of the PIMSN model on aspect-level sentiment analysis tasks.(4)A sentiment analysis system for user comments is designed and implemented.Based on the above research results,according to the data characteristics and scene requirements of sentiment analysis system,the algorithm proposed in this thesis is implemented.Users can log in to the comment sentiment analysis system,and the system can realize the functions of automatic extraction,analysis and prediction of comment sentences.