Research On Text Sentiment Analysis For Recommendation System

With the improvement of web 2.0,people have shifted from passively receiving information to actively participating in information dissemination,not only to obtain useful information from social media,e-commerce and other activities,but also to publish relevant comments on the Internet more conveniently and quickly.In view of the information redundancy and mixed praise and criticism of large amounts of data texts,in order to fully understand the thoughts and attitudes of the Internet users,text sentiment analysis based on artificial intelligence emerges.Currently,the research results of coarse-grained text sentiment analysis for document level and sentence level are becoming more and more mature,but they cannot cope with the fine-grained sentiment analysis task of multiple objects and multiple sentiment in the same text.At the same time,the recommendation system tasks,such as extracting key information from the complex content then scoring and predicting unknown items according to the user's intentions,have become research hotspots in the follow-up application of text sentiment analysis.Based on aspect-level of fine-grained text sentiment analysis,this thesis focuses on its strong dependence on external work and insufficient semantic extraction,and explores the use of deep learning to achieve end-to-end aspect-based text sentiment analysis,which integrates aspect detection and sentiment analysis to achieve more accurate and efficient deep-level text sentiment analysis.And sentiment analysis is added in the recommendation system to build a personalized recommendation system for user-item rating prediction.The main contents are as follows:1.Researching on the text sentiment analysis model based on BERT(Bidirectional Encoder Representations from Transformers).The BERT pre-training model is used for text preprocessing and related feature extraction,by combining the token,position and segment information to generate word vectors from the sentence level;through the multi-layer Transformers,necessary sequence information can be fully extracted to reduce the workload of the downstream network.Using softmax function as the classification network,model comparison and result analysis are performed on Laptop14 and Rest14 datasets,which verified the text processing advantages of BERT model in aspect-based text sentiment analysis.2.Studying on text sentiment analysis that integrates GRU(Gate Recurrent Unit)and attention mechanism.Based on the good results obtained by BERT feature representation,the downstream network adopts bidirectional GRU network and multi-head attention composed of self-attention mechanism to extract contextual information and global information respectively.With the complementary advantages,the fusion model fully learns high-level semantic information and further improves the precision of text sentiment analysis.The validity and domain adaptability of the model are verified on multiple publicly fine-grained text sentiment analysis datasets.3.Designing a recommendation system based on text sentiment analysis.Based on the traditional recommendation system,the aspect extraction and sentiment analysis of items are carried out through aspect-based text sentiment analysis,and the attention mechanism is used to extract the user's specific concern to the aspect needs.Combined with the aspect and sentiment matching degree between users and items,the unknown item is scored and predicted,realizing a personalized recommendation system.The corresponding Mean Square Error(MSE)and Mean Absolute Error(MAE)are reduced on multiple Amazon datasets,which further proves the effectiveness of the fine-grained text sentiment analysis model and the feasibility of the recommendation system based on text sentiment analysis.