Research On Tibetan Emotion Analysis Method Based On Deep Learning

With the rapid development of information technology in recent years,a large number of texts containing various opinions and emotions are distributed on various social platforms,and it is extremely important to conduct sentiment analysis research on such texts.Among the different languages,the research on sentiment analysis in Tibetan is still at a preliminary stage of development compared with Chinese and English.There are few research results related to sentiment analysis in Tibetan,and the research methods used by researchers are relatively backward,so it is urgent to promote the development of sentiment analysis research in Tibetan.The lack of publicly available and widely recognized datasets limits the research on Tibetan sentiment analysis,and most of the current research methods on Tibetan sentiment analysis are based on machine learning and a single deep learning model,so this paper addresses this part of the problem.This thesis draws on the results of Chinese and English sentiment analysis research,and uses deep learning methods,which are currently in full swing in the field of natural language processing,to study Tibetan sentiment analysis.In this thesis,four main areas of work are carried out for the research of Tibetan sentiment analysis based on deep learning methods,as follows.1.To address the lack of public datasets in the field of Tibetan sentiment analysis,a Tibetan sentiment analysis dataset has been constructed and made public on the Github platform for relevant researchers to use.The Tibetan sentiment analysis dataset was constructed by selecting 10,000 sentences containing two types of sentiment polarity from two publicly available Chinese sentiment analysis datasets,weibo＿senti＿100k and Chn Senti Corp,and then the original dataset was translated and proofread by members of the team with relevant expertise using sophisticated Chinese-Tibetan machine translation tools combined with manual work,and the final construction was completed.A Tibetan dataset for sentiment analysis.2.For Tibetan sentiment analysis methods are mostly based on a single deep learning model,each model has different advantages and disadvantages,so this thesis proposes a hybrid neural network-based model-AL-TCBAT applied to Tibetan sentiment analysis.The model uses the ALBERT pre-training model to generate word vectors,followed by local extraction of Tibetan text sequence features by Text CNN model in the same direction first to obtain relatively important text information.Subsequently,the extracted features through the Text CNN model are feature-built again by a bidirectional Bi LSTM layer to extract deeper information in them.Finally,the attention mechanism is incorporated to make the network pay more attention to the parts with larger weights among many features.The experimental results show that the model can improve the accuracy of the classification model.3.In response to the problem that single-channel hybrid neural networks can cause more serious loss of features as the number of network layers increases,this thesis proposes a multi-channel hybrid neural network model-AL-DCBAT,which first inputs the word vectors generated by the ALBERT pre-training model into the Text CNN and In order to further improve the extraction ability of key sentiment features in the Bi GRU part,an Attention mechanism is introduced after the Bi GRU model.The outputs of the two parts are fused and input to the normalization layer to obtain the final sentiment classification.The experimental results prove that the model can improve the classification accuracy of the model on the basis of reducing the depth of the lateral network.4.To address the problem that traditional deep learning models have limited ability to extract sequential text features and commonly use increasing model depth to improve the final classification effect,this thesis proposes an ALSGraph SAGE model based on graph neural network approach for Tibetan sentiment analysis.Firstly,the model is trained by ALBERT pre-training model to obtain the initial sentence feature vector.Second,to further improve the quality of the feature vector,this paper proposes the method of fusing sentiment word features,where each Tibetan sentence in the dataset is extracted with sentiment words and its features are randomly initialized by Embedding,and then the initial sentence features are fused with the sentiment word features as the final sentence features.Then,in the graph construction,the Tibetan sentences and tags are constructed as text-label graphs to complete the graph data construction.Finally,the graph data is input into Graph SAGE graph neural network model for feature learning and the final sentiment analysis results are obtained.The experiments prove that the proposed ALSGraph SAGE model achieves a good classification accuracy.