Introducing Persona Ecosystem Playground (PEP): Generating Behaviorally Grounded Personas from Think-Aloud Sessions

Abstract

Persona generation methods rely primarily on structured data like surveys and analytics, leaving rich qualitative sources such as think-aloud transcripts and interview records underutilized. We present Persona Ecosystem Playground (PEP), a system that uses LLMs with retrieval-augmented generation (RAG) to generate personas from unstructured qualitative data while maintaining systematic traceability to source evidence. Demonstrated on think-aloud transcripts from 67 participants in an educational chatbot study, PEP generates five personas with 77% of attributes directly supported by transcript evidence (cosine similarity ≥ 0.8), including 96% direct support for behavioral attributes and 90% for goal attributes. The integration of RAG-based grounding, cosine similarity validation, and Rao’s Quadratic Entropy diversity measurement into a single automated pipeline distinguishes PEP from prior persona systems that lack systematic traceability. This work establishes a validated persona generation foundation for multi-group stakeholder modeling, where distinct personas representing different interdependent groups can be generated from heterogeneous qualitative sources.