Abstract
Inductive reasoning is a core component of human intelligence. In the pastresearch of inductive reasoning within computer science, formal language isused as representations of knowledge (facts and rules, more specifically).However, formal language can cause systematic problems for inductive reasoningsuch as disability of handling raw input such as natural language,sensitiveness to mislabeled data, and incapacity to handle ambiguous input. Tothis end, we propose a new paradigm (task) for inductive reasoning, which is toinduce natural language rules from natural language facts, and create a datasettermed DEER containing 1.2k rule-fact pairs for the task, where rules and factsare written in natural language. New automatic metrics are also proposed andanalysed for the evaluation of this task. With DEER, we investigate a modernapproach for inductive reasoning where we use natural language asrepresentation for knowledge instead of formal language and use pretrainedlanguage models as "reasoners". Moreover, we provide the first andcomprehensive analysis of how well pretrained language models can inducenatural language rules from natural language facts. We also propose a newframework drawing insights from philosophy literature for this task, which weshow in the experiment section that surpasses baselines in both automatic andhuman evaluations. We discuss about our future perspectives for inductivereasoning in Section 7. Dataset and code are available athttps://github.com/ZonglinY/Inductive_Reasoning.