Sentiment Analysis Of Movie Reviews Based On Attention-Tree LSTM

In recent years,with the continuous development of the domestic film and television industry,the increase in people’s demand for spirituality,and the COVID-19 epidemic,China’s film industry has presented new trends and patterns.Nowadays,movies are not only a form of entertainment for people to spend time but also one of the most important symbols of a country’s cultural confidence.Due to the rapid development of the online industry,a large number of movie review websites and platforms have emerged.On these movie review websites and platforms,there are many reviews about a certain movie,and these reviews contain the subjective emotional tendency of the movie viewers towards the movie,which has a certain significance for the general audience to make correct movie viewing decisions.Given the difficulty of traditional lexicon and machine learning-based sentiment analysis techniques to effectively process the massive amount of text,this thesis uses deep learning techniques to conduct an in-depth study of movie reviews,which mainly includes the following two parts.First,the sentiment classification of movie reviews is done based on the Tree-LSTM(Tree-structured Long Short-Term Memory)model.The model includes syntactic syntax,which can accurately parse complex utterances and obtain more information from relatively distant nodes.The main work contains:using crawler technology to collect the review information of the influential "Hi,Mom" movie in 2021,and obtaining a high-quality corpus for sentiment analysis of movie reviews after pre-processing the data;performing descriptive statistics on keywords to mine deep information of movie reviews;using Borderline-SMOTE method to deal with the data imbalance;using Python’s The third-party module SnowNLP combined with manual annotation to perform sentiment annotation;LDA-based thematic analysis of positive and negative review sets;GloVe method is used to train and obtain word vectors suitable for the movie domain.Through experimental comparison with SVM,LSTM,Bi-LSTM,and other models,the Tree LSTM model is effective in sentiment classification,and the experiments show that the model is suitable for the movie domain.Second,to further optimize the model effect and solve the problem that the Tree LSTM model ignores the different focuses of sentence expressions,this thesis proposes a new method Attention-Tree LSTM,which uses the Tree LSTM model to extract semantics,locates,and weights each word,and then uses the attention mechanism to study the semantic features extracted to show the focus.To further validate the performance of the model,a public dataset with a larger amount of data is used in this thesis.For the dataset processing,the review data are classified into five categories:very poor,poor,average,recommended,and highly recommended based on the ratings.In the sentiment five classification experiments,the F1 value of Attention-Tree LSTM can reach 0.91372,which is 0.635%higher than the F1 value of Tree LSTM,and the rest of the evaluation indexes are also better than other comparable models,and the experiments show that the model has higher classification ability for movie reviews.