Research On Sentiment Information Extraction And Classification Based On Deep Learning

In recent years,the user number and coverage of mobile Internet have been developing rapidly.After completing online consumption,consumers usually publish some comments of goods or services on the service providing platform,this kind of comments often include multiple relative aspects of the target,and more fine-grained sentiment analysis has become an increasingly widespread requirement.Aspect based sentiment analysis is a main research direction of fine-grained sentiment classification.Related work is mostly divided into two types,one is based on the aspect.This kind of work requires a specified set of aspect categories,which results in poor scalability and flexibility in practical applications.The other type only extracts aspect terms or opinion terms and cannot obtain complete sentiment information.By analyzing the defects of the traditional method in dealing with opinion words,this thesis come up with a new ABSA task and a relative solution which could obtain more sentiment information,the specific work is as follows:(1)Aiming at the problem that some existing work ignores opinion terms in ABSA tasks,this thesis proposed a method of extracting sentiment information tuple based on deep learning model.The deep learning sequence annotation model is used to extract aspect words and opinion words,and the corresponding relationship between different words is judged by the classification model.In view of the lack of traditional methods in terms of word vector representation,a BERT-based extraction model of aspect words and evaluation words was proposed,and the performance of various downstream models in this task was studied through experiments.(2)This thesis proposes a task to completely extract aspect words and opinion words in the comment text and their corresponding sentiment tendencies,designs and implements a pipeline method solution based on the deep learning model for this task as well as studies the influence of different labeling strategies on the results.Aiming at the problems of error propagation and lack of context information found in the pipeline method,this thesis proposes a joint model based on BERT-BiLSTM underlying parameter sharing,which alleviates the problem of error propagation to a certain extent.Meanwhile,different tasks are generated synergistic interaction through sharing parameters.Finally,the experiment result proves that the model proposed in this thesis has a significant improvement compared with the pipeline model.(3)This thesis designed and implemented an aspect-based evaluation mining system,exploited the method proposed in this thesis to real hotel review data and achieved the effect of aspect-level sentiment analysis by extracting triples of sentiment information.