Bill-oriented Stance Prediction

From the "Asia Pacific Rebalancing Strategy" proposed by Barack Obama,to the "China–United States trade war" launched by Donald Trump,to the "Indo-Pacific Strategy" released by Joe Biden,the United States,as the world superpower,takes a series of measures to limit China’s development.Before the United States implements a series of measures,most of them need to be sponsored and voted by congressmen and then signed by the president.In recent years,The U.S.Congress has passed many bills that interfere with China’s sovereignty,such as "Taiwan Travel Act" and "Uyghur Human Rights Policy Act of 2020".Therefore,predicting the stances of congressmen on different bills has important strategic significance for China to formulate corresponding policies and safeguard national security.To solve the lack of dataset,the selection of features of congressmen and bills is too simple,and the relationship between congressmen is not fully considered,this thesis firstly constructs a high-quality dataset and then considers the understanding of political texts,feature selection,and the relationship between congressmen to construct a bill-oriented stance prediction model.The specific work of this thesis is as follows:(1)A bill-oriented stance prediction dataset based on multi-source data is constructed.To solve the lack of dataset in this research field,this thesis obtains multi-dimensional data from seven data sources,and constructs a bill-oriented stance prediction dataset,including congressman information dataset,bill information dataset,and voting information dataset.(2)An embedding vector model for the semantics of political texts is trained.Using billoriented stance prediction dataset,pretrain BERT-Base model released by Google to obtain Politics-BERT,which can better understand the semantic information of American political texts.Simultaneously,based on the understanding of the American political system and characteristics,the religious belief,ethnic group and political view of congressmen,the policy area and sponsor’s party of the bill are considered,and then using Politics-BERT to extract the feature vectors of congressmen and bills.(3)The model for predicting the stance of congressmen on bills is designed.To solve the defect that most studies ignore the relationship between congressmen or the construction of the relationship graph between congressmen is unreasonable,this thesis builds a political relation graph based on the party,state,and committee of congressmen and then using heterogeneous graph attention network to construct political relation graph model.Update the feature vectors of congressmen based on the relationship between congressmen.(4)The model performance is validated and the bill-oriented stance prediction system is implemented.Using bill-oriented stance prediction dataset,this thesis designs various experiments with other research models and variants of our model,which verifies the superiority of our model and the validity of each module of our model.At the same time,based on our model,the bill-oriented stance prediction system is designed and implemented.