The Research On Sentiment Analysis On Chinese Text

Social network platforms bursts with the rapid development of Internet. The users share their personal opinions and comments on products on the social network platforms daily. Using nature language proccess and machine learning techniques to analyze the mass user generated text data and mining the opinions towards specific events, person or products by users is becoming an important way in social opinion monitoring and product after-sales information feedback. Thus, the research on sentiment analysis in texts has shown great practical and commercial value.This study investigates the techniques on Chinese text sentiment analysis from two aspects, namely sentiment features extraction and transfer learning, respectively. Firstly, it is observed that the existing techniqens for sentiment features extraction and representation in sentiment analysis lacks of the method to integrate deep learning model and sentiment linguistics resources. Target to this problem, we propose a novel sentiment analysis appraoch which incorporates convolutional neural networks and word sentiment sequence features. This approach is based on word representations. By incorporating convolutional neural networks and the existing sentiment linguistics resources, the text data are mapped to vectors with sentiment features. The evaluations on Chinese Opinion Analysis Evaluation 2014(COAE 2014) dataset show that the proposed approach outperforms the baseline system. It achieves 0.97% and 1.58% F1 improvements on the positive opinion detection and the negative opinion detection, respectively. This approach provides a feasible scheme to incorporate deep learning model and the sentiment linguistic resources. It is a very possible new research direction in text sentiment classification.The domain correlation and sample selection bias problem are widely existed in sentiment data. Such problem affects the construction of the optimal classification model. Thus, from the view of constructing feasible training data, which is fit to the distribution of testing, by selecting the quality train samples from train data, this study investigtes three transfer learning based approaches, namely k-nearest neighbors based, classifier iteration based and knowledge transfer based on deep Gaussian processes, respectively. The evaluations on COAE 2014 dataset show that all of the three approaches outperform the baseline system. In whchi, the knowledge transfer learning approach based on deep Gaussian processes acheives the best performance. It achieves 5.01% and 2.94% F1 improvements on the positive opinion detection and the negative opinion detection, respectively. It is shown that this appraoch reduces the negative impacts introduced by the text data domain correlation and sample selection bias problem effectively.