E-commerce Commentary Sentiment Analysis System Based On Deep Learning

The analysis of emotional polarity of Chinese phrases in complex contexts has always been a problem in the field of machine learning.The emergence of artificial neural networks has opened a window to solve this problem.Sentiment analysis has also become a hotspot in the research.This thesis aims to solve this problem by introducing Attention mechanism to Temporal Convolutional Nets（TCN）model.This thesis chooses Chinese commodity comment phrase on E-commerce platform as research object to automatically analyze and judge the polarity（positive and negative）of each comment phrase,and get comprehensive evaluation of the goods.In order to analyze the polarity of phrase,word to vector mechanism is introduced to map text information to distributed representation.This thesis chooses some mature recurrent neural networks models（LSTM,Bi-directional LSTM,LSTM Attention）and Temporal Convolutional Nets（TCN）model to train and test using E-commerce comment dataset.By comparing and analyzing the test results of these models,this thesis find that LSTM Attention model is better than those models without Attention mechanism（LSTM,Bi-directional LSTM and TCN）.And for the models without Attention mechanism,the performance of TCN model is better than other models（LSTM and Bi-directional LSTM）.So the idea of combining TCN model with Attention mechanism is proposed.The test results shows that TCN Attention model is better than the previous models on E-commerce comment dataset,which proves the correctness and innovation of combining TCN model with Attention mechanism.Secondly,this thesis also analyzes the misidentification problem of the model and further improve the performance of the model by extending the dataset.In order to promote sentiment analysis model to practical,this thesis presents the imple-mentation of E-commerce commentary sentiment analysis system based on deep learning,which based on Keras+Tensorfow machine learning platform.The E-commerce comment sentiment analysis system includes two parts:offline model training subsystem and online inference sub-system.The overall architecture and main functional modules of the system have been detailed and designed in this thesis.This thesis has studied various neural networks models of convolutional neural networks and recurrent neural networks.In-depth analysis of problems encountered in data collection,data preprocessing,model design and model optimization.Which makes TCN Attention model shows superior performance on Chinese E-commerce comment dataset.Finally,the TCN Attention model proposed in this thesis achieves accuracy:94.94%;"Positive"classification precision:94.08%,recall:92.30%,F₁-Measure:93.18%;"Negative"classification precision:93.66%,recall:93.90%,F₁-Measure:93.77%on test set,which has a significant improvement comparing with other models.At last,this thesis forecasts the broad prospect of the application of TCN Attention model in E-commerce comment system,and point out the problems of TCN Attention model and the idea of optimization.