Abstract
The interdisciplinary subject of computational social science is revolutionized by the explosion of the big data and the scientist’s ability to analyze the them. In this thesis, we approach several computational social science problems, namely blame tie extraction for the financial crisis (political science), sustainable development goal tracking (environmental science), and topic modeling for COVID-19 pandemic (public health). We approach these problems with state-of-the-art NLP models, and achieve promising results. In our first work, we study “who blames whom” for the 2008-2010 U.S. financial crisis. We design a new task called “Blame Tie Extraction”, and develop BiLSTM-based supervised models to extract blame relations from thousands of news articles. We further enhance the model performance with the pretrained language model ELMo. Secondly, we explore the use of pretrained masked language models to track the progress of the sustainable development goals. Specifically, BERT is used to synthesize evidence from Wikipedia in a self-supervised way. We find that pretrained language models can be used to gain insights into the progress of sustainability goals. In our third work, we examine what people are talking about on the social media platform Twitter during the COVID-19 pandemic. We use unsupervised LDA-based topic modeling to detect topics and conduct a comprehensive analysis of the distribution and dynamics of the topic characteristics for several typical countries.