Research On Aspect Based Sentiment Analysis Based On Customer Reviews

The industry competition pattern is showing a white-hot trend,so merchant management becomes more difficult,wider and more refined.At the same time,customers are pursuing more perfect service quality and product quality in the industry.Analyzing and discussing the massive customer reviews and tap the application value contained in them is meaningful.By analyzing customer reviews,it is possible to describe consumption feelings and reflect the quality of merchants’ services.Therefore,extracting value information from reviews is beneficial to both merchants and consumers.Traditional sentiment analysis focuses on coarse-grained sentiment analysis at the chapter level and sentence level,and analyzes the text as a whole to obtain the overall sentiment polarity.If there are multiple evaluations in the text and the corresponding emotional polarities are inconsistent,the effect often cannot cater to individuals.Aspect Based Sentiment Analysis（ABSA）aims to extract the user’s emotional bias towards a certain aspect or attribute of a product,which can provide users with more accurate information.Aiming at the problem of Aspect Based Sentiment Analysis,this article studies and designs an effective deep learning algorithm model according to Aspect Term Sentiment Analysis（ATSA）and Aspect Category Sentiment Analysis（ACSA）.The main contents and achievements of the article are as follows.1.For the ATSA task,This thesis proposes a hybrid network model of Bi LSTM-Gate CNN based on ALBERT,a lightweight text representation.The ALBERT pre-training model is used for text representation,and the Adapter module is added,which greatly reduces the training time and the amount of learning parameters of vector representation by taking advantage of the light weight.In the process of text feature extraction,the Bi LSTM network is combined with the gated convolutional network,the model focus on local features consequently and context-related information at the same time,and enhances the context learning ability of the model.The comparison experiment with the current mainstream model shows that the model has a great improvement in the evaluation index.2.For the ACSA task,it includes three sub-tasks: aspect word recognition,opinion extraction,and aspect category and sentiment polarity.For the aspect word recognition task,the existing models have the problems of complex structure and slow learning speed.This thesis proposes a Bi GRU-CRF aspect word recognition algorithm based on ALBERT,which improves the training time overhead by reducing the model learning parameters,while ensuring the performance.For the task of opinion extraction,referring to the question-and-answer mechanism and incorporating prior knowledge with aspect,a Ro BERTa-QA text representation-based training method is proposed to enhance the model’s ability to capture opinion information.For the task of aspect category and sentiment polarity,this thesis design an improved multi-class multi-task model based on Ro BERTa and multi-head attention mechanism.By conducting comparative experiments on public datasets,the results verify that the model has a good effect on sentiment classification tasks.3.An aspect based sentiment analysis system based on customer reviews is designed and developed.The aspect based sentiment analysis model put forward is applied to practical problems,and the results are visualized and displayed to users to verifie the pertinence and usefulness of the model proposed in this thesis.