Research On Deep Learning And Transfer Learning In Chinese Sentiment Classification

Text sentiment classification plays an important role in natural language processing research.In industry,the emotional classification of product reviews serves as one of the user's market feedback,which is of great significance to manufacturers and sellers.Compared to English,the Chinese text is more complex.In this regard,our paper mainly focuses on Chinese product reviews and explores the emotional classification of texts.The recurrent neural network extracts and fuses the information of different time sequence in the text,to realizes the interpretation of the text's semantics,emotional tendencies and other characteristics.When the text sequence is too long,due to the loss of information in the process of transmission,the traditional recurrent neural network can not always extract the overall text information;at the same time,when the text's emotional representation different at begin and end,the model has a chance to misjudgment of the overall emotional representation.To solve these problems,our paper proposes a bidirectional GRU model for Chinese text sentiment classification.Experiments show that this model can more accurately obtain the emotional representation of long time series texts by combining the gating mechanism and the two-way word processing method of text.In addition,our paper explores the influence of Dropout strategy on Chinese text sentiment classification.The results show that proper Dropout model can effectively prevent overfitting and improve the generalization ability of the model in text sentiment classification.In machine learning,people often use the method of transfer learning to improve the model in response to the lack of datasets in specific areas.Our paper proposes a model-based transfer learning method to realize cross-domain learning in Chinese text sentiment classification tasks.Specifically,first use the bidirectional GRU model to train other domain datasets and save model weights.Then for a specific domain model,import the previous weights and retrain the data to update some of the model parameters so as to achieve cross-domain model migration.The analysis shows that transfer learning can effectively improve the generalization performance of the same model in different areas of learning.At the same time,our paper verifies the existence of noise in sentiment classification data and the impact of noise on transfer learning.