The PhD project aims to bridge the connection between two areas that have long been developed separately, causal inference and natural language processing (NLP). Specifically, we aim at the following goals: (1) methodological improvements of NLP models using causal inference theories, and (2) applying causality+NLP to important social applications. For the first goal, we apply the following causal inference research to build more robust NLP models: independent causal mechanisms, invariant causal predictions, disentanglement, algorithmic fairness, and causal discovery by minimum description length. For the second goal, we look at social problems such as the causality among policies, public opinion, and media spread of information. We use NLP models to extract features from large-scale online text such as social media, and news articles, apply causal inference techniques such as do-calculus, and involve interdisciplinary knowledge from political science such as policy responsiveness and democracy political theory.