Causal maps, consisting of labeled nodes and directed edges, capture individual mental models, which can be aggregated to form complex group-level maps. While aggregate maps leverage distributed expertise, they often lose transparency, creating barriers for community members. Software solutions have attempted to address this issue, but diagrammatic representations can still hinder accessibility. Large Language Models (LLMs) have been proposed to automatically explain conceptual models as text, but they lost information and often focused on sentence generation whereas we need reports formed of paragraphs. In this paper, we design, implement, and evaluate a method that automatically translates causal maps into coherent narratives without model loss. Our evaluation compares 24 community detection algorithms to decompose maps into topical communities and utilizing 3 LLMs to generate coherent paragraphs. The optimal combination produces text with readability levels comparable to human-written paragraphs. Our method is open-source, allowing modeling groups to adapt it for their stakeholders.