What Makes it Ok to Set a Fire? Iterative Self-distillation of Contexts and Rationales for Disambiguating Defeasible Social and Moral Situations

TL;DR

This paper introduces a novel iterative self-distillation method to generate high-quality, diverse contextual explanations for moral judgments, enhancing the understanding of nuanced human moral reasoning.

Contribution

It presents a new approach combining self-distillation, filtering, and imitation learning to improve the validity and diversity of moral reasoning contexts and rationales.

Findings

Produced a dataset of 1.2 million contextualized moral judgments

Achieved high human agreement rates of 85.9% to 99.8%

Final model outperforms intermediate models significantly

Abstract

Moral or ethical judgments rely heavily on the specific contexts in which they occur. Understanding varying shades of defeasible contextualizations (i.e., additional information that strengthens or attenuates the moral acceptability of an action) is critical to accurately represent the subtlety and intricacy of grounded human moral judgment in real-life scenarios. We introduce defeasible moral reasoning: a task to provide grounded contexts that make an action more or less morally acceptable, along with commonsense rationales that justify the reasoning. To elicit high-quality task data, we take an iterative self-distillation approach that starts from a small amount of unstructured seed knowledge from GPT-3 and then alternates between (1) self-distillation from student models; (2) targeted filtering with a critic model trained by human judgment (to boost validity) and NLI (to boost diversity); (3) self-imitation learning (to amplify the desired data quality). This process yields a student model that produces defeasible contexts with improved validity, diversity, and defeasibility. From this model we distill a high-quality dataset, \delta-Rules-of-Thumb, of 1.2M entries of contextualizations and rationales for 115K defeasible moral actions rated highly by human annotators 85.9% to 99.8% of the time. Using \delta-RoT we obtain a final student model that wins over all intermediate student models by a notable margin.