Research On Sentiment Analysis Method Of Commodity Reviews Based On Deep Learning

With the full coverage of the Internet in the e-commerce industry,various business service platforms have developed rapidly,various types of data information are carried by these platforms.Compared with audio-visual multimedia data,text data in the Internet consumes less resources and is easy to be transmitted on the Internet,so text information is the most widely expressed form.Text sentiment analysis in natural language processing studies how to find useful information from text.Due to the limitations of computing power and optimization algorithm,there are many problems in the current commonly used sentiment analysis algorithm,such as more parameters,only works well for specific data sets and real-time performance cannot meet the needs of practical application.In this thesis,the online reviews text data of e-commerce platform is taken as the research object,the BiLSTM algorithm is selected as the basic algorithm,combining attention mechanism and part of speech,the semantic information in text is learned from a large number of reviews data samples,then the text features and sentiment classification results are obtained,so as to achieve the purpose of accurate extraction of sentiment in text data.The deep learning model and knowledge distillation method are combined to realize the accuracy and lightweight of sentiment analysis model and optimize the model performance.According to the designed classification model,the commodity reviews sentiment analysis system based on deep neural network is studied,the system can automatically predict the emotional state expressed by the text content according to the user's operation,so as to help people mine the text information and make correct judgments.The main work of this thesis is summarized as follows:(1)Research and improvement of sentiment analysis algorithm for commodity reviews based on deep learning.This thesis uses the BiLSTM algorithm as the starting point to design a pos-BiLSTM-Att optimization method.Through the fusion of the attention mechanism,the main feature is given a higher attention probability value,thereby improving the processing speed and accuracy,and then improving the text classification effect.It also embeds part-of-speech information to better learn the characteristics of review sentences.In order to make the designed pos-BiLSTM-Att model not only have a higher accuracy of sentiment analysis,but also keep smaller complexity.In this thesis,the knowledge of ALBERT-FN teacher model is transferred to pos-BiLSTM-Att student model by knowledge distillation method,so as to obtain the pos-BiLSTM-Att model with superior performance.(2)The performance test and analysis of sentiment analysis algorithm for commodity reviews based on deep learning.The experimental environment and experimental data are introduced,and the core algorithm of this thesis is compared with other sentiment analysis algorithms from the prediction accuracy and real-time response effect of the algorithm.Experiments show that the three key indicators of the pos-BiLSTM-Att model proposed in this thesis are higher than those of other comparative models on different datasets.In addition,the distilled pos-BiLSTM-Att model can achieve better prediction results with smaller amount of parameters and faster response time.(3)Design and implementation of sentiment analysis system for commodity reviews based on deep learning.In this thesis,Word2vec is selected as the word vector model,and the distilled pos-BiLSTM-Att model is selected as the system sentiment analysis model.And the overall design scheme of the system is constructed according to the system design requirements and application scenarios.The basic functional modules,framework and workflow of the system are designed,the system registration and login as well as the functions of each sub module are realized respectively,which proves that the system can meet the expected requirements.