Enhancing Dynamic Profiles on Voting Advice Applications Using Social Media Analysis

Voting advice applications are web-based tools designed to guide citizens toward the parties and candidates close to their political preferences. However, they require the generation of candidates' profiles, and they must fill out a questionnaire scientists have designed. Since this procedure is time-consuming and expensive, we propose a data-driven model which can facilitate the construction of the profiles using the recent developments in the Natural Language Processing domain. To this end, we first gather the politicians' Twitter data and YouTube videos. Then we clean the data and translate the tweets and video transcripts. We aim to use a question-answering (QA) model, which takes a question and a context and predicts the answer using the context. Our Twitter and YouTube data is significant. Therefore it is not immediately suitable for QA models. According to the topics of the questions in VAAs, we create shorter context vectors using manually defined keywords. Then we apply QA models to the tweet and YouTube data. Our experiments show that this approach can automatically find correct answers in the data and help human annotators.