Research On Sentimental Orientation Classification For Public Opinion Text

The rapid development of the Internet has allowed more people to express their emotions,attitudes,and opinions on the Internet,so a large number of public opinion texts have been formed.How to mine and analyze the public's sentimental attitude from the mass of public opinion texts,help guide the direction of public opinion,and provide support for public opinion monitoring and early warning has become a more popular research direction at present.Sentimental orientation classification as an effective method for processing public opinion texts has also received widespread attention from scholars.At present,deep learning methods have been widely used in sentimental orientation classification tasks,and the introduction of attention mechanisms has further improved the accuracy of classification.However,most of the existing methods that integrate attention mechanisms only focus on the learning of word semantic information,and pay less attention to the learning of part-of-speech information,ignore the sentimental connection between words and parts-of-speech,and there is still room for improvement in classification accuracy.In addition,the data imbalance problem that commonly exists in actual scenes will interfere with the training of the model,thereby affecting the effect of sentiment classification.Therefore,the following researches are carried out:(1)Aiming at the problem of data imbalance,Focal Loss is introduced into sentimental orientation classification.By dynamically adjusting the weight of various samples,the contribution of the samples in a larger number of classes to the loss function is reduced,which makes the model pay more attention to the learning of the samples in a small number of classes.(2)Aiming at the problem of insufficient utilization of the part-of-speech information in the existing deep learning methods,Part-of-Speech based Transformer Attention Network(pos-TAN)is proposed.This model uses Transformer Attention Network(TAN)as the basic network structure,and uses the self-attention mechanism to represent text features,and integrates part-of-speech information into different levels of the network,combining POS-Attention to learn rich sentimental information in part-of-speech.(3)In order to keep the pos-TAN model lightweight and still have high classification accuracy,this thesis draws on the knowledge distillation,and uses the pre-trained BERT model as a teacher network to guide the training of the pos-TAN model.It makes that the pos-TAN model will learn the parameters and generalization capabilities of the BERT model.(4)The prototype design and implementation of the sentimental orientation classification system for public opinion texts.The sentimental orientation classification technology is applied to actual data to predict and analyze the sentiment tendency of public opinion texts.The experimental results show that the introduction of Focal Loss improves the classification accuracy of the TAN model on each dataset by 0.5% to 0.8%,and the fusion of part-of-speech information further improves it by about 1%.Moreover,the pos-TAN model achieved the best classification effect on the five datasets in the comparison experiment with the baselines.In addition,the distilled pos-TAN model can achieve a similar BERT prediction effect with a smaller amount of parameters and faster inference speed.