Leveraging multidimensional features for policy opinion sentiment prediction

Abstract

Previous online policy opinion analyses based on social media data have focused on topic detection and sentiment classification of policy opinion after a given period following policy implementation. These approaches are limited and inefficient because they provide no opportunity to change citizens’ opinions once they have been formed. Furthermore, incorporating auxiliary information to enrich semantic representations is vital and challenging due to limited texts, and a lack of both semantic information and strict syntactic structure. Therefore, we propose a novel framework to extract and integrate multidimensional features from user-related and policy-related social media information and predict policy comment polarity in the policy release phase. First, we construct four machine learning models for model-induced features to capture topic-related and opinion-related features and identify the policy-opinion nexus. In addition, we integrate basic and behavioral user features. Then, we leverage multidimensional features to construct a stacked learning model for predicting the policy opinion. Finally, we conduct experiments on 20 policy comment datasets to demonstrate that our prediction framework can effectively predict public opinion about a policy once it is released. Our model provides key insights into policy opinions in advance and can enable policymakers to engage in better policy communication before opinion formation.