A unified evaluation framework for revising inappropriate arguments with edit-level guidance

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In an appropriately-phrased argument, the language supports credibility and emotional engagement while remaining proportional to the issue at hand. Arguments that fall short of these properties undermine public deliberation and the quality of democratic discourse, motivating research on their automatic revision. Because ex-post moderation is costly and time-consuming, suggesting revisions at creation time is a promising alternative. We unify two prior approaches to revising inappropriate arguments under a common edit-operation framework: a PPO-based approach that aligns a large language model (LLM) to generate full rewrites from which edit suggestions are extracted ex-post using a diff tool, and a GRPO-based approach that makes the LLM generate self-contained span-level edit suggestions directly. The framework defines the desired properties of the task and introduces two additional quality criteria: pattern conformity at the argument level and appropriateness at the edit level. Placing both paradigms on a common footing yields an approach-agnostic basis for evaluating edit suggestions, independent of whether they are generated directly or derived from a full rewrite. The results of a moderator-style user study, in which human moderators iteratively selected and applied edit suggestions from our approaches, suggest that overall revision quality is high (79.2% of arguments revised to full appropriateness), while also revealing span-level shortcomings in rewriting-derived edit suggestions that motivate the quality criteria introduced here. A comprehensive evaluation of eight argument revision approaches reveals that approaches generating edit suggestions directly and optimizing jointly over multiple quality criteria achieve human-likeness competitive with large prompted models, while approaches optimizing for appropriateness alone systematically produce edit suggestions that violate structural quality requirements. Across both studies, the two generation paradigms emerge as failing in different ways: edit suggestions extracted from full rewrites are frequently interdependent and lose fluency when applied selectively, whereas edit suggestions generated directly are self-contained but individually less effective and, hence, need iterative application. Together, these contributions establish argument revision via span-level edit suggestions as a unified and interpretable task, with direct implications for human-in-the-loop moderation and AI-assisted deliberation.