Research and Application of Cross-domain Text Sentiment Classification Based on Deep Learning

Abstract

With the rapid development of information technology, the popularity of the Internet has become higher and higher. The Internet is full of subjective reviews from netizens about various products or things. These contents are produced in large numbers every day, and they actually have far-reaching value and research significance. User reviews often have subjective sentimental tendencies, and sentimental analysis of these data can help researchers understand user feedback in time. Recent academic achievements have shown that the neural network model can be constructed by the deep learning theory, and the trained model can well realize the task of text sentiment classification. This task based on deep learning requires a large amount of data with sentimental labels as a training set to obtain better effects and avoid over-fitting. It is well known that the cost of obtaining manually labeled data is high, and some new domains often have poor performance of text sentiment classification due to lack of enough labeled data. Therefore, the research on cross-domain text sentiment classification is imminent. The main goal of this thesis is to use the technology of artificial neural network and transfer learning to provide an effective solution to this problem, and to improve the generalization ability of the neural network model in classification, achieving the task of cross-domain text sentiment classification in the absence of datasets. This thesis has completed the following three aspects:

Due to the existence of many parameters in the deep learning model, it is necessary to use a large-scale labeled dataset as the training set to ensure the accuracy of text sentiment classification. However, the authoritative and mature dataset is limited, and the cost of constructing a dataset is expensive. In addition to adopting the public dataset, this thesis also uses the technology of web crawler to get the reviews of several popular products on the Steam website as the experiment. source data. These textual data contain a total of 20,000 reviews with user tags. After being cleaned, filtered, and pre-processed, they will be organized into a new dataset.

For the text sentiment classification tasks based on deep learning, this thesis establishes and designs the system flow according to the experimental requirements. In this thesis, the text in the dataset will be vectorized by the GLOVE tool firstly, and then it will be imported into the embedded layer in the neural network model built for the experiment. All processed data will be inputted into the system for training while the model will be evaluated in a cross-validation manner. After comparing and analyzing several groups of models using different neural networks, this thesis finally adopted the structure of BI-LSTM. This model can effectively improve the accuracy of text sentiment classification to 92.7%, and its effect on classification is significantly better than other models.

In the cross-domain sentiment classification task, this thesis proposes a model-based transfer learning method according to the relevant theory of transfer learning, which realizes the shared learning between similar but different domains. The transfer learning system built in this experiment will first train the source domain dataset through the BI-LSTM model, and save the structure and weight parameters of the model. After that, the new model for the target domain will import the weight of the source domain, freeze the weight of some network layers, and retrain the data of the target domain to update some parameters. The final analysis of contrast experiments shows that the transfer learning method can effectively improve the generalization ability of the neural network model when cross-domain learning. To some degree, transfer learning can reduce the dependence of deep learning on large-scale datasets, thereby reducing the consumption of computing resources and saving the training time.