Automated Conspiracy Theory Detection and Narrative Consensus Tracking in Social Media

A longstanding grand-challenge problem in AI is to build machines that are able to think and interact like humans do. A specific embodiment of this problem is a generalization of the cocktail party problem encountered in signal processing and blind signal separation: If an AI agent were to drop in at a crowded cocktail party then can it separate out and reconstruct the different underlying stories and narratives being discussed from a mixture of fragments of all the on-going conversations. Such a problem has taken on a renewed urgency: Narratives play a defining role in influencing critical decisions and worldviews of both the society at large and individuals, but the continual emergence of a multitude of conflicting narratives –enabled by large-scale adoption of social media– has created a global emergency, where the basic tenets of civil society and governance are being increasingly compromised. These narratives, some of which can be labeled as conspiracy theories, are composed of numerous characters connected by semantically diverse relationships situated in multiple and overlapping contexts. Injecting false facts happens in the context of such discussions, and solving such a misinformation problem is beyond a supervised classification task in natural language processing (NLP). In this dissertation, we develop a pipeline of interlocking computational and statistical modules - based on NLP tools and complex network theories– to extract meaningful narrative networks by distilling millions of social media posts. We develop a framework for semantic parsing of such narrative graphs (e.g. who are outsiders, their motivations and threats, and strategies of insiders) and evaluate the quality of these automatically derived communities in different ways. This framework takes a step towards enabling human-smart AI, where it can view the world as a human would