The famous saying, “Common sense is not so common,” by Voltaire (Dictionnaire philosophique, 1764), depends upon human experiences and individual perceptions. Commonsense reasoning (CR) in artificial intelligence (AI) includes domains like natural language processing (NLP), computer vision, and robotic manipulation. This paper discusses CR and knowledge in AI.
The authors categorize successes in automated CR into four types: taxonomic reasoning, temporal reasoning, action and change, and qualitative reasoning. Taxonomic reasoning defines three basic relations: an individual is an instance of a category; one category is a subset of another; and two categories are disjoint. The temporal reasoning automates knowledge and reasoning about time, duration, and time intervals. The authors describe the theory of action, events, and change by simplifying constraints: events are atomic, single actor, and perfect knowledge. They further provide extensions to domains, including continuous domains, simultaneous events, probabilistic events, multiple agent domains, imperfect knowledge domains, and decision theory. According to the authors, “qualitative reasoning is about the direction of change in interrelated quantities.”
The authors classify five challenges in automating CR: virtually untouched or partial understanding of the domains, varying logical complexity in different situations, plausible reasoning involvement, long-tail phenomenon, and difficulty in discerning the proper level of abstraction. They further comprise the objectives in CR research, such as reasoning architecture, plausible inference, range of reasoning modes, painstaking analysis of fundamental domains, breadth, independence of experts, applications, and cognitive modeling. The authors divide reasoning techniques into three types: crowdsourcing, web mining, and knowledge-based approaches, which further include mathematically grounded, informal knowledge-based, and large-scale approaches.
The authors doubt that the problems of CR will be easily solved. However, they recommend the creation of benchmarks, an evaluation of the program CYC, the integration of various AI approaches, the inclusion of alternative modes of reasoning in mainstream approaches, and a better understanding of human CR. This paper is an interesting read for those who are working in the area of CR and commonsense knowledge in AI applications.