Barbara Gendron 

Date/Time: 25.02.2026, 12:00 

Location: D2.0.030 

Abstract 

This work investigates ontology-driven control of LLM-based conversational agents. Key conversational properties are defined in an ontology, complemented by a rule-based strategy specifying when agent utterances must satisfy predefined constraints. An initial LLM–ontology hybrid architecture enabled systematic control of English CEFR levels [1] through a simple strategy [2]. Since then, characteristics other than language level alone have been represented in the ontology, now covering three aspects of control. This allows us to test different fine-tuning scenarios by controlling a single aspect at a time (single-aspect) or several aspects (multi-aspect). To achieve conversational control that has a greater impact on LLM generation, I explored fine-tuning procedures based on reinforcement learning, typically using GRPO algorithm [3]. The latter has significantly improved results in single-aspect controlled generation, while multi-aspect generation control appears to be a much more demanding task. This ontology-driven conversational control has been implemented in a conversational agent for the experimental use case of job interview practise. As part of my visit to DPKM, I am extending this framework to a complex real-world application: the development of a conversational agent to support ontology engineering. This involves regulating both the factual content and the stylistic dimensions of the interaction, thereby providing a substantial extension of the current control setup.

Bio 

Barbara is a third-year PhD student working in MosAIk team at LORIA, under the supervision of Mathieu d’Aquin and Gaël Guibon. She works on improving conversational models based on Large Language Models (LLMs) through the building of a conversation-dedicated ontology. She graduated from École des Mines de Nancy (French engineering school, Master level) and hold a Master’s degree in Data Science from University of Luxembourg. Her main research interests revolve around Natural Language Processing. She is interested in the development of machine learning models for knowledge extraction and natural language processing, as well as in the applications of Data Science in life sciences and medicine. As an example of her previous research projects, she worked on meta-learning methods for emotion prediction in conversational context during her Master Thesis [4].

[1] https://www.coe.int/en/web/common-european-framework-reference-languages/level-descriptions

[2] Barbara Gendron, Gaël Guibon, Mathieu D’aquin. Towards Ontology-Based Descriptions of Conversations with Qualitatively-Defined Concepts. TOTh International Conference, 2025.

[3] Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y.K., Wu, Y., Guo, D.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models (2024).

[4] https://b-gendron.github.io/projects/projects-1/