Explainable Commonsense AI

With the stakes of AI increasing, there is a recognition that a key requirement of a human-centric AI is explainability. This recognition has inspired a range of methods for making AI explainable, either during or after its main inference process. Such explainable AI (XAI) efforts form an extensive taxonomy of approaches differing in their scope, stage, input/output format, result, and functioning. The most popular idea in these XAI methods is to localize the part of the input or the network that is most responsible for a given prediction. As much of the real-world inference relies on implicit commonsense information, it requires specialized XAI methods that explain a decision by including this implicit information in their explanations. This chapter discusses the challenges in developing commonsense explanation systems, including the implicit nature of common sense, alignment between the explanation and the model result, and incompleteness of explanations. Then, we describe several neuro-symbolic methods that can be mapped to popular XAI functioning categories: structure leveraging explanations through path generation, architecture modification model based on compositional reasoning, and example-based explanations via case-based reasoning. The chapter also covers another category of commonsense explanations, based on language modeling mechanisms, enriched with rationale models and chain-of-thought reasoning. The chapter concludes with a summary of the state-of-the-art explainable commonsense AI and a discussion of its limitations and open challenges. We discuss customary procedures and metrics for evaluating commonsense explanations, challenges with the quality of the current commonsense explanations, and gaps between commonsense explanations in AI and social science.