Sentiment Analysis Based On Deep Learning Ensemble Model

With the rapid development of the Internet and the explosive growth of Internet users,online shopping has gradually integrated into people's lives.Especially in recent years,online shopping has even become a way of life.In 2018,about 71% of Internet users are online shoppers.Because such a large number of online shoppers shop online,the comment text data on various e-commerce platforms has exploded.Most of these comment text data contain the user's opinions and emotional tendencies.Emotional analysis has gradually become an important research goal in the field of natural language processing.With the explosive growth of data and the rapid development of computer technology,traditional statistical research methods can no longer meet the demands of today's big data analysis.Therefore,this paper uses deep neural networks combined with ensemble learning methods to analyze emotions in texts which collected from major e-commerce platforms.At first,this paper introduces the traditional research methods used in the past for emotional analysis research,and introduces the mainstream methods in sentiment analysis research from the aspects of statistical language models,word embedding and deep learning models,and then fasttext word embedding with time series characteristics is used to replace the word2 vec word embedding in text's vectorization,it well solves the problem of insufficient expression of text temporal features in the past.Next,the emotional features of the corresponding words are constructed through the emotional dictionary,and finally a series of comparative experiments are conducted on the sentiment analysis of the comment texts by combining the neural networks and ensemble learning.Experiments show that in the analysis of sentiment analysis,using fasttext word embedding to represent text,the GRU model obtained by training is slightly better than the GRU model trained by using word2 vec word embedding.In the experiment,whether using fasttext word embedding or word2 vec word embedding,the classification accuracy of the GRU model is about 3% higher than CNN model.For the problem of insufficient expression in one-way circulation GRU model,the Bi-GRU model is used for sentiment analysis,and the classification accuracy of the test set reaches 91.14%.In addition,this paper combines the Bagging method in ensemble learning to train multiple Bi-GRU-based learners for ensemble model.The final result has a classification accuracy of 93.09% on the test set,which is about 2% higher than unensemble model in accuracy rate.