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

1. Problem Statement

Large Language Models (LLMs) have revolutionized human-computer interaction but suffer from a "black-box" nature, leading to issues with predictability and lack of personalization. While LLMs are powerful, they struggle to consistently satisfy complex, multi-dimensional constraints (e.g., specific language proficiency levels or emotional polarity) required for goal-oriented domains like healthcare, customer support, and HR. Existing constrained generation methods are often:

• Costly: Requiring full fine-tuning or complex hybrid architectures.
• Model-dependent: Tied to specific model architectures.
• Inconsistent: Struggling to maintain constraints over long conversations or across multiple dimensions simultaneously.

The authors aim to solve this by proposing a model-agnostic, lightweight framework that uses ontologies to define conversational strategies and control LLM outputs without altering the underlying model architecture.

2. Methodology

The proposed framework integrates structured knowledge (ontologies) with LLMs through a three-stage process: Definition, Strategy, and Fine-tuning.

A. Ontological Definition & Strategy

• Descriptors: Key conversational aspects are modeled as descriptors (e.g., language complexity, emotional load, polarity).
• Ontology Construction: These descriptors are formalized within an ontology (built in Protégé) using description logic. This allows for structured, logic-based reasoning.
• Conversation Strategy: A strategy is defined as a set of rules governing how descriptor values evolve during a conversation. The system uses an ontology reasoner (Pellet) to infer the required class for the next utterance based on the conversation history and user input.

B. Data Preparation & Labeling

• Label Wrapping: To train the LLM to adhere to these constraints, the authors use a "label-wrapping" technique. Utterances are enclosed in brackets with their corresponding ontology class: [o.C : Ci] utterance [o.C : Ci].
  • Pre-utterance labels guide the generation.
  • Post-utterance labels ensure alignment and control.
• Datasets: The framework utilizes multiple datasets (DailyDialog, GoEmotions, CEFR-S, etc.) to create training data for specific tasks.

C. Fine-Tuning Procedure

• Causal Language Modeling (CLM): The LLM is fine-tuned using a lightweight LoRA (Low-Rank Adaptation) approach. This significantly reduces trainable parameters, making the process efficient and applicable to smaller models.
• Model Agnosticism: The method is applied to seven state-of-the-art open-weight models (Llama 3, Qwen, Phi, Mistral, DeepSeek-R1).

D. Evaluation

• Zero-Shot Generation: The model is prompted with a user query and the left-hand bracketed label (the target constraint). The model generates text, and the right-hand label is stripped during post-processing.
• Metrics:
  • Classification Metrics: Accuracy and Weighted F1-score to measure if the generated text matches the target ontology class.
  • Task-Specific Metrics: Mean Absolute Error (MAE) for proficiency levels (ordinal data) and Matthews Correlation Coefficient (MCC) for polarity.
  • Semantic Similarity: BERT-F1 Ratio (Br) to ensure the control mechanism does not degrade the semantic quality of the output.

3. Key Contributions

1. Novel Ontology-Driven Framework: A method to control LLM generation using ontological definitions and reasoning, enabling modular and explainable control strategies.
2. Lightweight & Model-Agnostic: The use of LoRA adapters and a generic labeling strategy allows the framework to be applied to various open-weight LLMs without heavy computational costs.
3. Two Distinct Use-Cases:
   • Proficiency-Level Control: Adapts the agent's English level (CEFR A1–C2) based on the user's detected proficiency, using a "harder-only" strategy (the conversation complexity cannot decrease).
   • Polarity Profile Control: Manages the emotional load (loaded vs. non-loaded) and polarity (positive, negative, neutral) of responses to encourage critical thinking and debate while mitigating toxicity.
4. Reproducibility: The authors publicly released code, data, and ontologies.

4. Results

The framework was evaluated on seven LLMs across the two use-cases.

• Quantitative Performance:
  • Fine-tuned models consistently outperformed pre-trained baselines in adhering to constraints.
  • Proficiency Control: Fine-tuning significantly improved the distinction between CEFR levels (e.g., separating simple vs. complex sentences based on FKGL indices). The best models achieved F1 scores around 0.24–0.31 compared to 0.06–0.14 for baselines.
  • Polarity Control: Fine-tuned models showed improved ability to generate specific polarities (e.g., generating negative content when requested). DeepSeek-R1-8B achieved the highest F1 (0.44) and MCC (0.40) in this task.
  • Class Coverage: Unlike pre-trained models which often failed to generate certain classes (e.g., specific CEFR levels), fine-tuned models successfully generated content across all target classes.
• Qualitative Observations:
  • Generated text showed qualitative differences in vocabulary and sentence structure corresponding to the requested proficiency levels.
  • The framework successfully demonstrated the ability to shift conversation tone (e.g., from positive to negative to encourage debate) without losing coherence.
• Semantic Quality: The BERT-F1 Ratio (Br) indicated that while the generation changed to meet constraints, the semantic similarity to the original model's output remained relatively high, suggesting the control did not severely degrade fluency or coherence.

5. Significance and Future Work

• Significance: This work demonstrates that structured knowledge (ontologies) can effectively guide LLMs to produce predictable, personalized, and safe outputs. It bridges the gap between the flexibility of LLMs and the rigidity required for domain-specific applications. The approach is particularly valuable for applications requiring strict adherence to guidelines (e.g., medical advice, educational tutoring) where hallucination or uncontrolled tone is unacceptable.
• Limitations:
  • Quantitative results still have room for improvement, suggesting the CLM fine-tuning may not fully internalize complex ontology concepts.
  • The evaluation relies heavily on classifiers; human evaluation is needed to assess true conversational engagement.
  • The concepts modeled (like "emotional load") are inherently subjective.
• Future Directions:
  • Exploring Reinforcement Learning (e.g., PPO, DPO) to use ontology compliance as a reward signal.
  • Broadening the framework to more complex, real-world conversational strategies.
  • Conducting full human evaluations to assess user experience and perceived compliance.

In conclusion, the paper presents a robust, lightweight, and interpretable method for controlling LLM conversations, proving that ontological reasoning combined with efficient fine-tuning can significantly enhance the controllability of generative AI.