Towards Intelligent Agents that Learn to Reason and Communicate

We propose to explore how to create intelligent agents that learn by reading, in order to perform commonsense reasoning. Commonsense reasoning is crucial for intelligent systems: It is part of the shared background in working with human partners, and provides a foundation for future learning. It requires massive amounts of knowledge, far beyond the scale of what can be encoded by hand. Thus learning by reading is a crucial step towards creating such broadly capable agents. Our hypotheses are: (1) qualitative representations are a crucial part of commonsense knowledge and (2) analogical reasoning and learning provides robustness in reasoning, plus human-like learning of complex relational structures. Unlike deep learning systems, for example, analogical learning systems can handle relational structures such as arguments, proofs, and plans, while learning with orders of magnitude less data. This research should help pave the way for intelligent systems that can interact with, and learn from, people using natural modalities, as well as make progress on understanding the nature of human cognition. Using the Companion cognitive architecture, we propose to explore the following ideas: 1. Knowledge Integration. We will explore how systems should integrate new knowledge into what they already know, scrutinizing it for inconsistencies and implications, as people seem to do. The accuracy of knowledge will be estimated by using it to reason and tracking the correctness of answers to identify where problems are. Re-reading to find alternate interpretations will also be explored, as a means of improving the accuracy of subsequent reading as well. 2. Learning to Communicate Better. Our recent progress on learning constructions via analogical generalization suggests that integrating analogy deeper into semantic interpretation could help make natural language systems that are more adaptable and flexible. We will explore co-learning of language and reasoning, using question/answer pairs to derive what interpretations should be drawn from a question and what patterns of plausible inference should be used in constructing answers. We will explore the role of analogy in natural language generation, essentially inverting the mappings between syntax and meaning learned in constructions, to communicate a system’s explanations to its human partners. Potential Naval Relevance: Many tasks relevant to US Navy operations could potentially benefit from intelligent systems that can reason, communicate, and learn robustly. These include teaching and tutoring, operations planning, decision-support systems, and software assistance for developing complex scientific, engineering, and policy models.