Conversational Control with Ontologies for Large Language Models: A Lightweight Framework for Constrained Generation

Abstract: Conversational agents based on LLMs have recently emerged as powerful tools for human-computer interaction. Nevertheless, their black-box nature implies challenges in predictability and a lack of personalization, both of which can be addressed by controlled generation. This work proposes an end-to-end method to obtain modular and explainable control over LLM outputs through ontological definitions of aspects related to the conversation. Key aspects are modeled and used as constraints; we then further fine-tune the LLM to generate content accordingly. To validate our approach, we explore two tasks that tackle two key conversational aspects: the English proficiency level and the polarity profile of the content. Using a hybrid fine-tuning procedure on seven state-of-the-art, open-weight conversational LLMs, we show that our method consistently outperforms pre-trained baselines, even on smaller models. Beyond quantitative gains, the framework remains model-agnostic, lightweight, and interpretable, enabling reusable control strategies that can be extended to new domains and interaction goals. This approach enhances alignment with strategy instructions and demonstrates the effectiveness of ontology-driven control in conversational systems.

• The paper introduces a lightweight framework combining ontological descriptors with LoRA fine-tuning to enable controlled and interpretable conversational generation in LLMs.
• It operationalizes conversation strategies through CEFR-based language proficiency and sentiment polarity, achieving improved compliance metrics like F1 and MAE.
• The framework is model-agnostic, data-efficient, and reusable, providing practical applications for enterprise, educational, and emotionally sensitive domains.

Ontology-Guided Conversational Control in LLMs: Framework and Empirical Insights

Introduction

The black-box nature of LLMs presents ongoing challenges for controlled and predictable conversational agent behavior, particularly in domains requiring adherence to user-specific interaction strategies or constraints. "Conversational Control with Ontologies for LLMs: A Lightweight Framework for Constrained Generation" (2604.04450) presents a modular, interpretable methodology for controlled text generation in LLMs using ontology-grounded descriptors and strategies. By formulating key conversational aspects as ontological entities and operationalizing these via lightweight LoRA-based fine-tuning, the framework achieves explainable and reusable content control, demonstrated on proficiency and sentiment/polarity axes.

Framework and Methodology

The proposed approach consists of a two-level structure: (1) ontological definition of conversational descriptors and strategies, and (2) causal LLM (CLM) fine-tuning to enforce these constraints.

Descriptor and Strategy Formalization: Conversation-specific aspects—such as language proficiency (mapped to CEFR levels) and polarity/emotional load—are identified as descriptors. These are formalized within a machine-interpretable ontology using description logics, enabling both static (utterance-level) and dynamic (conversation-level) constraint specifications. Strategies governing descriptor evolution across conversation turns are also ontologically instantiated, facilitating explicit, logic-based reasoning for next-utterance class selection.

Generation Control: LLMs are fine-tuned on datasets where each utterance is bracket-labeled with ontological class information, providing both pre- and post-utterance cues to encourage consistent conditional generation. Multiple open-weight LLMs (Llama3, Qwen, Phi-3.5, Mistral, DeepSeek-R1) are fine-tuned using LoRA adapters, ensuring model-agnostic and resource-efficient transfer.

Evaluation Protocols: Controlled generation is primarily evaluated in a zero-shot fashion by prompting the model with a user input plus a control class indicator in bracketed form; the output is then reclassified by the relevant ontology-based classifier (decision tree or RoBERTa), and compliance is measured using accuracy, F1, MAE (for ordinal CEFR compliance), and MCC (for polarity profile alignment).

Use Cases and Experimental Design

Proficiency-Level Control

The agent's response complexity is governed by the highest CEFR level detected in the user's previous utterances (the “expressed-is-understood” paradigm, enforcing monotonic complexity). Descriptor rules for CEFR levels are explicitly induced via a decision tree from linguistic features (Flesch-Kincaid Index, Gunning-Fog, lexical diversity, etc.) and formalized as ontology classes. Fine-tuning data is sourced from CEFR-labeled corpora and DailyDialog.

Polarity Profile Control

Each utterance is classified as emotionally loaded/non-loaded and positive/neutral/negative, yielding six atomic classes. Strategies are designed for debate engagement while reducing sycophancy: the agent inverses polarity to encourage critical thinking except for loaded negative/emotional statements, promoting productive disagreement and emotional awareness. RoBERTa-based classifiers handle descriptor annotation on dialogue corpora.

Results and Quantitative Analysis

The ontology-guided fine-tuning paradigm yields systematic improvements over pre-trained baselines in both control and compliance:

• Proficiency control F1 scores improved from near-zero (0.06–0.14 in raw models) to up to 0.31 for Llama3-8BF, with MAE reductions over 1 point and marked FKGL separation between CEFR classes.
• Polarity profile control shows similar gains, with fine-tuned models producing meaningful class-based differentiation (e.g., up to 0.44 F1 for DeepSeek-R1-8BF, 0.35 for Qwen2.5-7BF).

An important finding is that all classes are generable post-fine-tuning, unlike with unsupervised baselines (which exhibit catastrophic forgetting or class starvation). There are, however, model size/performance tradeoffs and variation in transferability across LLM families.

Qualitative analysis further reveals distinct lexical and structural changes aligned with control targets, although sycophancy and classifier misalignment persist as limiting factors in nuanced scenarios.

Semantic similarity measures (BERT-F1 ratio) indicate that content semantics (for proficiency) remain comparable pre- and post-control, validating the non-destructive nature of constraint imposition.

Practical and Theoretical Implications

Practically, the framework is model-agnostic and data-efficient, with LoRA enabling scalable deployment on both large and small LLMs and ontologies capturing reusable domain knowledge. The architecture is robust for enterprise, educational, and emotionally-sensitive domains where compliance and explainability are critical.

Theoretically, the results demonstrate that knowledge formalization via ontologies can significantly enhance LLM controllability, elevating beyond prompt-engineering or ad-hoc control tokens to logic-aware, interpretable conditional generation. This opens avenues for more composable, declarative control strategies, and facilitates integration with broader KB/ontology-driven AI systems.

Limitations and Future Work

While substantial gains are observed, fine-tuning in the CLM framework is not always sufficient for complete compliance, suggesting that integration with RL-based (e.g., PPO, DPO) protocols—with appropriate stability controls—may be required for more granular or higher-stakes applications. Furthermore, concept subjectivity (especially in emotional/polarity space) highlights the ongoing challenge of robust ontology engineering.

A crucial limitation remains the absence of comprehensive human evaluations to supplement classifier-based metrics. Conversational coherence, context maintenance, and perceived usefulness require real-user studies to validate practical efficacy.

Ethical risks include potential misuse for manipulation, requiring alignment safeguards and ethical oversight for domain deployment.

Conclusion

The integration of ontological descriptors and strategies with lightweight, LoRA-based LLM fine-tuning presents a rigorous, extensible framework for controlled conversational generation. The methodology unifies knowledge-driven reasoning with the expressive power of modern LLMs, achieving significant advances in predictable, explainable agent design. Extensions in RL-based training, dynamic prompting, and expanded ontological coverage are natural next steps to realize adaptive, ethically-aligned conversational systems at scale.