Sentiment Analysis Of Commodity Reviews Based On Deep Neural Network

The emotional trend of commodity reviews directly reflects the popularity of commodities in the market,and sentiment analysis can provide reliable analytical data for company sales.The complexity of the way of text expression leads to the unsatisfactory result of feature extraction by using traditional machine learning,and the fields of NLP requires a better models to get useful result.The main work of the thesis are as follows:1.Conversion to Chinese Characters into Digital Expressions and correcting the problem of polysemy by a multi-semantic word training method.Extracting subject words based on the TF-IDF method to provides input support for subtasks.Experiments have shown that these methods can obtain more efficient digital text information.2.This paper proposes a fine-grained Attention mechanism.In the model,Attention feature extraction is mainly performed on the text level,the sentence level and the paragraph level.The experimental results show that the model effectively improves the accuracy of sentiment classification.3.In massive data,the model is easy to convergence in local optimum and it takes a long time to spend on model training.In order to solve these problems,a migration mechanism based on Fine-tuning is adopted.The trained and mature model parameters are migrated to the current model for iterative training,and the learning rate algorithm is improved based on the STLR algorithm,so that the model can converge quickly.4.The comment text has various characteristics,such as multiple important information and complex features.In this paper,a parameter sharing network based on Sluice is built for multi-task feature sharing.Firstly,a multi-task model is built on different data sets,and then the parameters learned by the hidden layer are shared by blocks,so that the model can be learned less obviously.Or feature information that is ignored.Combining with the method mentioned above,this paper proposes a deep neural network model based on TS-BiGRU.Experiments show that the model can effectively improve the training speed and prediction results in the emotional classification task.