CRII: III: Robust and Explainable AI Agents with Common Sense

点击翻译按钮获取中文摘要

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2). This project will gain an understanding of how to create Artificial Intelligence (AI) agents that provide commonsense explanations about real-world narratives. Current AI agents lack commonsense mechanisms to explain their judgment of everyday stories and they cannot be applied to novel scenarios. This award will enable AI agents to reason in novel situations and to explain their decisions. The project will focus on two key aspects of stories: understanding situations and judging the adequacy of actions in context. The project will test the ability of AI agents to complete narratives and to provide commonsense explanations on the task of explainable natural language inference. The explainability of AI agents can be expected to improve public trust in AI technologies. Robust and explainable AI with common sense is also critically missing in social AI assistants that aim to increase the participation of children with Autism Spectrum Disorder and the elderly with Alzheimer's dementia. The investigator will design a new set of lectures and a full course on the topic of “AI assistants with common sense”, which will be taught both at USC as well as internationally. Interdisciplinary research will be facilitated via summer internships, and participation in the existing University of Southern California (USC) Center for Knowledge-Powered Interdisciplinary Data Science and NSF Research Experiences for Undergraduates programs. The investigator will partner with USC's Center for Engineering Diversity and Women in Science and Engineering, in order to recruit members of historically underrepresented groups for research on this project. The investigator will partner with USC's K-12 STEM Center to engage K-12 students from historically underrepresented groups.This award will create a paradigm shift in the development of AI agents, by combining advances in neural language modeling with high-level explanations based on logical axioms and commonsense knowledge. State-of-the-art technology is not adequate for this goal: neural methods cannot infer causal links between events and the motivations and goals of the agents directly from narratives, whereas commonsense axioms and knowledge resources alone cannot handle the contextual variations in human language. The team of researchers will build AI agents that use common sense to explain their reasoning. To do so, the researchers will leverage commonsense knowledge and axioms about agent psychology and event causality in order to enrich story corpora. The enriched data will be used to pre-train neuro-symbolic agents to complete open-world narratives and justify their completion with commonsense explanations. The researchers will measure the impact of representative techniques, axiomatic theories, and knowledge dimensions on understanding narratives about situations and actions.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.