Research And Application Of Text-based Sentiment Analysis Technology

With the advent of 5G mobile network and the rapid development of portable mobile terminal network devices,people can get a lot of information from the Internet,and express their opinions on shopping websites and social media anytime and anywhere,which result in the explosion of the text data with individual user sentiment.The task of sentiment analysis based on text data has attracted much attention since it appeared,and has been widely studied in many practical applications.The purpose of this task is to automatically detect the sentimental polarity containing in the text,explore the user's views on products,hot events,problems,things,etc.,and help individuals or organizations make better decisions.In this thesis,the task of sentiment analysis of text with different granularity is studied,and extensive experiments are carried out on the algorithm.The main research results are as follows:(1)This thesis proposes a multi granularity attention mechanism based MAHG model for document-level sentiment analysis tasks,which applies bidirectional gated recurrent unit to establish hierarchical structure,capture long-distance context semantic information,construct cross layer connection to make up for information loss caused by deep-seated model,and propose a multi granularity attention mechanism to obtain more effective context feature representation,so as to identify sentiment more accurately.Based on this mechanism,we study the MAHG model based on bidirectional encoder representation from transformers(BERT)and use Bert to generate text embedded information,which further improves the performance of the model.The experimental results show that MAHG model and BERT-MAHG model both achieve good performance on document-level sentimental data sets.(2)For sentence-level sentiment analysis task,this thesis proposes a lightweight end-toend CSRDS model based on dual channels,which can improve the performance of the model and ensure the training efficiency as much as possible.The model deploys spatial coding channels and temporal coding channels to model text in-depth from two dimensions of time and space,obtains the local phrase vector representation and the global context vector representation for integration,and designs a double positive mechanism focusing on salient features.CSRDS model constructs more comprehensive sentimental features to solve the problem of text sparseness in multi granularity level of sentence-level sentiment analysis tasks,and reduces the computational complexity.The effectiveness of the proposed model and the superiority of training efficiency are verified by comparing the proposed model with several baseline models on four public English sentence-level sentiment data sets.(3)The aspect-level sentiment analysis task aims to discover the sentiment tendency of users in different aspects in sentences containing multiple aspects,which is different from the previous two kinds of sentiment analysis oriented to the whole text.This thesis proposes an aspect-level sentiment analysis model of RPISC based on information flow selection mechanism.This model uses multi-channel one-dimensional convolutional neural network as the core architecture to obtain multi-scale local features,which is different from most of the current aspect-level sentiment analysis models based on recurrent neural network variants such as long short-term memory network.In addition,we fully consider aspect information and introduce relative position information to capture the association information of aspect words and their corresponding sentimental opinion words,and design information flow selection mechanism to extract the sentimental features of specific aspects.After experimental verification,compared with the baseline model,the model has obvious advantages in aspectlevel sentiment analysis tasks.(4)For aspect-level sentiment analysis tasks,this thesis proposes a TAMG sentiment analysis model based on tripartite attention mechanism.In order to break the limitation that it is difficult to connect aspect words and corresponding sentimental opinion words in a given natural language word order,this thesis introduces syntactic information into the model and use dependency syntax tree to represent sentence structure.Then,a graph convolution network is constructed to capture the semantic information better in the dependency syntax tree.A tripartite attention mechanism is proposed to infer and learn among aspects,context hidden layer representation and node feature representation,so as to better judge the information that is more important for specific sentimental orientation.Through extensive experiments,it is proved that the performance of the model is improved in many aspects.