Silverbullet or Mystery Box – LLM-based Soft Skill Classification in Volunteering

Volunteering is a vital pillar of critical infrastructures (CIs) and sustainable development goals (SDGs), fostering, e.g, civil protection or rescue/health/social services. Whether supporting CIs or SDGs, skill-based volunteering is key. Standardized knowledge about skills viable or necessary for certain volunteering opportunities is beneficial in the pre-engagement phase to enable effective skill use as well as in the post-engagement phase to leverage skill gain. It is unclear, however, in how far existing skill classification approaches - currently solely focusing on job postings on the labor market - can handle the nuanced, taskdriven, and prose-like descriptions of predominantly soft skills typical in volunteering opportunities. This paper addresses this gap by presenting a comparison of existing skill classification approaches, initially developed for labor market job postings, in the context of volunteering opportunities. Based on that, we propose using cache- and retrieval-augmented generation techniques for soft skill classification in volunteering, avoiding the high costs of LLM fine-tuning common in current methods. The effectiveness of these lightweight techniques is evaluated both quantitatively and qualitatively using a novel soft skill dataset with expert-labeled volunteering opportunities from a global volunteering platform.