Comparative Study of Sentient Anlysis using Modified Pre-trained Deep Learning Models for Chinese App Market Comment Texts

This paper compares sentiment analysis of Chinese user comment text of mobile applications based on both sentiment dictionary and machine learning/deep learning methods. The dataset is unique because of the inclusion of emoji, memes, slangs, “modernized Chinese,” etc, which are not found in other dataset or dictionary. The evaluation of the sentiment dictionary method shows that the prediction results are consistent with the sentiment polarity of the comments. The prediction results based on the fine-tuned and modified BERT pre-trained model achieve a precision rate of 94.8%. After converting the emoji to text and retraining, the precision rate of the algorithm improved to 95.2%.