Text Sentiment Analysis Based On Deep Learning Methods

With the rise of Web 2.0, more and more data can be found on the Internet. People produce texts information via writing blogs, microblogs, product reviews, film reviews and so on, which are mainly unstructured data contain sentiments or opinions of the writer. If the sentiment beneath all the data can be properly extracted, there will be a great improvement for automatic decision support, risk analysis for cyberspace public opinion, information alert and commodity sales, which is of great significance for scientific research and practical application.Tranditional methods for sentiment classfication can be categorized as knowledge-based methods, machine learning based methods and hybrid methods. When the data is small and the senmantics are not divese, these methods are efficient for text sentiment analysis. However, facing more and more data with abundant representation methods, trainditional methods are no more functioning well. There is an urge for a brand new methods.Since 2006, deep learning has drown attention of a lot of people from industry and academia. Although the architecture of deep learning based methods are quite similar to traditional neural netwok, with the totally different data representation and training methods, it is possible to train rather efficient models with more hidden layers. Deep learning based methods has became the state-of-art solution for video recognition, speech recognition and so on.Convolutional Neural Network and Recurrent Neural Network are to efficient model in deep learning, the former is ablet to extract local feature and the latter is designed for learning general sequences. It is hard to achieve satisfied results in text sentiment analysis with any single methods, which lead to the emergence of a hybrid model concatenating these two models. This paper propose three modifications for such hybrid model, which are optimizing the input vector sequence into a fixed length list, designing a new activation function to alleviate gradient dimishing and improve generalization ability and applying Maxpooling technique to extract the max value of local feature. The modifications are proved to be efficient from the results obtained from the experiments on Yelp 2015 dataset. Besides, in order to investigate the influence of important parameters of the model, we conduct several contrast experiments.