Research And Implementation Of Aspect Level Sentiment Classification Based On Multiple Feature Fusion

The Internet has penetrated into all aspects of people’s life.People deeply rely on various network platforms for clothing,food,housing and transportation,while freely express their views and opinions.These contents are mainly presented in the form of text and contain a lot of information.Consumers use the information to obtain an objective review of businesses and make consumption decisions.Businesses can use the information to guide the improvement of goods.It is very difficult to obtain the required information from a large number of texts.These texts can be analyzed automatically by using sentiment analysis algorithm.However,when making decisions,businesses and consumers not only focus on the overall emotional tendency of comments,but also the performance of goods in different aspects.Therefore,this thesis focuses on the practical need and studies the aspect level sentiment classification task based on user comments.There are many sub-tasks of aspect level sentiment classification task.This thesis focuses on the more challenging sub-tasks of implicit aspect word detection and implicit aspect sentiment analysis.Implicit aspect words can’t be extracted directly from sentences,and the model needs to have a certain ability of assimilation and abstraction.With the development of neural network technology,the algorithm of aspect level emotion classification task is also constantly updated.Convolutional neural network,Recurrent neural network and other technologies are applied.However,there are still many difficulties in this field,which are difficult to be solved by a single network model.This thesis focuses on the fusion of multiple feature expression models,solving the aspect level sentiment classification task,and develops a user comment oriented analysis system that can be applied in practice.The specific work content is carried out as follows:1.There are overlapping expressions of multiple aspect words in user comment text,which is difficult to be solved by current models.This thesis focuses on the fusion of Recurrent neural network,Capsule network,pretrained language model,and Graph neural network in feature extraction.In order to avoid the problem that a single model can’t learn all the useful text information,a network integrating multiple feature extraction models is designed.2.The current aspect level emotion classification model is relatively simple for the encoding of aspect words,which is usually randomly initialized.The embedding of aspect words does not contain any additional information but serves as the distinction between different aspect words.Considering the linguistic phenomenon that there may be an influence relationship between the emotional tendencies of different aspects in a comment,this thesis studies a coding method to strengthen the expression of aspect words.3.The upstream task of aspect category sentiment analysis is aspect category detection,and the correlation between the two tasks is very high.At present,most researchers study the two tasks separately,which will lead to the spread of errors.This thesis uses multi-task learning to process aspect categories detection and aspect categories sentiment classification simultaneously,and design a new loss function that can fully train both tasks at the same time.4.Design and develop an aspect sentiment analysis system for user reviews.The system mainly analyzes user reviews in the catering industry,supports two functions of aspect categories detection and aspect categories sentiment classification,and visualizes the analysis results.In addition,it also includes user login and historical data query functions.