Learning by Reading for Robust Reasoning in Intelligent Agents

We propose to use qualitative representations and analogical processing to explore how to create intelligent agents that can learn by reading in order to perform robust common sense reasoning. Common sense reasoning is crucial for many kinds of intelligent systems. Common sense reasoning rests on a foundation of broad knowledge about the world, beyond the scale of what can be hand-coded. Thus progress in common sense reasoning ultimately requires endowing intelligent agents with modalities for learning about the world that are expressive, scalable, and practical. A key modality is natural language, because language is expressive enough to describe principles, examples, and reasoning strategies while being natural for people. The ability to understand language is crucial for intelligent agents that interact with people or are trained by people. Moreover, such intelligent agents must be adaptable, learning on the job about how to improve their performance and how to better communicate with people. Learning by reading, using simplified English texts, offers a “sweet spot” which facilitates research on techniques to enable intelligent agents to understand language more deeply. It supports research on common sense reasoning, but also is a step along the way to theories that capture human-level natural language understanding.