Learning causality from textual data

Опубликовано 17 августа 2016, 2:35

It has been a long time quest of artificial intelligence to develop systems that can emulate human reasoning. Fundamental capabilities of such intelligent behavior are the abilities to understand causality and to predict. Those are essential for many artificial intelligence tasks that rely on human common-sense reasoning, such as decision making, planning, question-answering, inferring user intentions and responses. Much of the causal knowledge that helps humans understand the world is recorded in texts that express people's beliefs and intuitions. The World Wide Web encapsulates much of our human knowledge through news archives and encyclopedias. This knowledge can serve as the basis for performing true human-like prediction - with the ability to learn, understand language, and possess intuitions and general world knowledge. In this talk I will present Pundit - a learning system, which given an event, represented in natural language, predicts a possible future event it can cause. During its training, we constructed a semantically-structured causality graph of 30 million fact nodes connected by more than one billion edges, based on 150 year old news archive crawled from the web. We devised a machine learning algorithm that infers causality based on this graph. Using common-sense ontologies, it generalizes the events it observes, and thus able to reason about completely new events. We empirically evaluate our system on the 2010 news, and compare our predictions to human predictions. The results indicate that our system predicts similarly to the way humans do.