Human Evaluation

This page hosts a human evaluation study for the conversational agent I built during my PhD at LORIA, Controlled Conversational Models through Conversation-Dedicated Ontology, supervised by Mathieu d’Aquin (LORIA, CNRS) and Gaël Guibon (LIPN, Université Sorbonne Paris Nord).

The idea: guide the model with a conversation strategy defined from formal concepts and rules, giving a set of explicit constraints laid on top of a language model to make its behaviour more predictable and relevant to the use-case. The prototype simulates a job interview across five stages:

Phase 1 — Greetings and small talk2 turns
Phase 2 — Background and experience3 turns
Phase 3 — Technical discussion3 turns
Phase 4 — Short debate2 turns
Phase 5 — Summary and send-off2 turns

At each stage, the model is expected to respect constraints on language level, polarity, emotional tone, and fit with the current stage.

Taking part

For each context, you’ll see two possible responses and pick the one that fits best. You can also flag inconsistencies — the agent stuck in the wrong stage, contradicting itself, or forgetting something said earlier. No AI background needed, just read the guide first:

Questions, or want to take part? Get in touch.