Research On Stance Detection Method For Social Media

With the implementation of the “Belt and Road” initiative,exchanges between China and ASEAN countries are also increasing,and the analysis of the online public opinion of ASEAN countries has important theoretical significance and application prospects for improving the real-time monitoring and prediction capabilities of hot events in the region.Social media provides rich data resources for public opinion research,and researchers use machine learning,deep learning and topic models to carry out stance detection tasks,and propose a large number of effective research methods,especially pre-trained language models that have achieved excellent performance in stance detection tasks.However,this type of model is large-scale and requires a lot of computing resources and high time cost in practical applications.In view of the above problems,this paper focuses on the Philippines and conducts indepth research on stance detection methods.The main work of the thesis is as follows:(1)Most of the existing stance analysis datasets are oriented to political or social hot topics in Europe and the United States,and the analysis datasets for ASEAN public opinion are relatively scarce.In response to this situation,this paper constructed an English-language stance analysis dataset for the 2022 Philippine presidential election.In addition,two public datasets were used as experimental benchmark datasets to verify the effectiveness of the related model proposed in this paper.(2)Machine learning methods based on feature engineering rely on manual labor when constructing features and cannot be universal in the face of complex tasks,while pre-trained language models are large-scale and require a lot of computing resources and high time costs for practical applications.To solve the above problems,this paper proposed a stance detection model based on hybrid deep neural networks.The experimental results show that the macro average F-score on the SemEval-2016 Task6A dataset is 72.43%,which is higher than the best result based on traditional machine learning and lower than the best result on the pre-trained model of 75.1%,but greatly reduces the time cost and computing resources.In addition,it is found that the selection of FastText as the word embedding model can be used to jointly improve the efficiency of text stance detection.(3)Most of the existing stance detection work directly models the words in the text,ignoring the role of external knowledge in the representation of the text.To solve this problem,this paper proposed a stance detection model based on Wikipedia knowledge injection.The experimental results show that the F-score of the model is 85.9%,which is an improvement of 1.85% compared with BERT-Base.It demonstrates that external knowledge effectively enhances the performance of stance detection.The two stance detection methods proposed in this study are not only applicable to social media-oriented stance detection,but can also be extended to text classification tasks such as rumor detection,sentiment analysis,and fake news detection.