TLDR: This research introduces an ontology-based approach to control conversational AI by quantitatively defining qualitative aspects of conversations, such as user language proficiency. It uses linguistic descriptors and a Decision Tree Classifier to create formal rules for these features, which are then integrated into an ontology. This framework allows for fine-tuning Large Language Models to generate more predictable and personalized responses, demonstrated through a use-case of adapting an agent’s language level to a user’s English proficiency.
Conversational agents powered by Large Language Models (LLMs) have become the go-to technology for dialogue systems, thanks to their ability to generate fluent responses across many topics. However, their “black-box” nature means their outputs aren’t always predictable, sometimes leading to inaccurate information or even hallucinations. Moreover, these models often lack personalization, as they aren’t explicitly instructed on specific user interaction styles or personalities.
To build greater user trust and ensure accuracy, researchers are increasingly looking into generation control. This approach not only offers a path to predictable personalization but also improves the overall quality of generated content. Inspired by neurosymbolic AI, which combines symbolic reasoning with neural networks, and the integration of LLMs with Knowledge Graphs, a new study explores how generation control can be formally defined using ontologies.
The core of this research focuses on enabling conversational control by leveraging logical definitions of conversation characteristics, established within a dedicated ontology. Many important user-oriented aspects, such as emotion, conviction, satisfaction, or language level, are often defined qualitatively. These qualitative features, while crucial for understanding user-agent interactions, lack formal, quantitative definitions that can be directly integrated into ontologies for controlling conversations.
To address this, the proposed approach involves creating quantitative definitions for these qualitatively-defined conversational features. This is done by selecting specific pre-defined descriptors. Once these descriptors are identified, rules for subclass membership are extracted for each feature based on descriptor values. These rules are then expressed in description logic, allowing them to be integrated into an ontology. This integration provides benefits like consistency checking and inference capabilities. By incorporating this ontological information, after a reasoning step, into prompts during an LLM fine-tuning process, the researchers aim to control generation based on the defined conversation features.
A key use-case explored in the paper is “Proficiency-Level Control.” This involves modeling a user’s English proficiency and subsequently adapting the agent’s language level to match. The Common European Framework of Reference for Languages (CEFR) is used as the basis for defining proficiency levels (A1, A2, B1, B2, C1, C2), which are typically described using subjective qualitative statements. The challenge lies in translating these qualitative descriptions into consistent, factually-defined (quantitative) definitions that can be used in an ontological framework.
For feature selection, the study considers various linguistic features that describe text complexity. This includes well-known readability metrics like Gunning-Fog, Flesch-Kincaid, and Dale-Chall scores. Additionally, a wide range of linguistic properties are included, such as lexical features (e.g., named entity count, average word length), morphological and syntactic properties (e.g., coordination, conjunctions, subordinations), and discourse features related to rhetoric and connectivity (e.g., pronoun density, presence of indirect speech). All these features are either binary or numerical scores.
In their experiments, the researchers applied their framework to the Proficiency-Level Control use-case, defining `ProficiencyLevel` with sub-concepts from A1 to C2. They trained a Decision Tree Classifier (DTC) to establish the relationship between numerical descriptor values and qualitative proficiency levels. The DTC was chosen for its ability to provide explicit belonging rules for each class and to facilitate feature selection. The model was trained on the CEFR-T dataset, derived from the CNN/DailyMail dataset, which contains expert-annotated news texts, and validated using CEFR-SP, a dataset of CEFR-labeled single sentences. The model achieved an overall accuracy of 0.66, which is considered promising given the inherent subjectivity and inter-annotator agreement score of 0.41 for such concepts. The Mean Absolute Error (MAE) was 0.42, indicating that most errors were misclassifications between neighboring proficiency levels.
The explicit rules derived from the Decision Tree Classifier were then translated into description logic syntax and integrated into an ontology using Protégé. This allowed for the formal definition of concepts like `Utterance` with sub-concepts such as `A1LevelUtterance` up to `C2LevelUtterance`.
Compared to state-of-the-art deep learning approaches for CEFR-level classification, which often achieve higher accuracies (e.g., 0.75 with RNNs, 0.95 with BERT), this research highlights a crucial difference. Deep learning models are black-boxes, meaning they don’t provide the formal, explicit rules needed for ontological control. Furthermore, many existing deep learning studies rely on datasets like EFCAMDAT, which have been noted for label inconsistencies when mapped to CEFR levels and significant class imbalance.
Looking ahead, the research aims to use these DTC-inferred ontological definitions to annotate other datasets, which can then be used to fine-tune language models. A fine-tuning procedure, relying on Causal Language Modeling with the Llama3-8B-Instruct model and LoRA adapters, has already shown promising results. This fine-tuning process helps the language model learn to generate content according to the specified ontological concepts. For instance, the fine-tuned model demonstrated a better distinction of the Flesch-Kincaid index (a key readability metric) across different CEFR levels compared to the raw model.
Also Read:
- ProToM: An AI That Understands and Encourages Helpful Behavior Among Independent Agents
- Sticker-TTS: A New Framework for Smarter AI Reasoning Through Historical Experience
In conclusion, this work presents an innovative ontology-based approach to control conversational generation by quantitatively defining qualitative conversation aspects. By using linguistic descriptors and a decision tree classifier, the study creates structured definitions that ensure consistency and transparency within an ontological framework. The success in modeling CEFR levels for proficiency control demonstrates the method’s effectiveness in handling subjective concepts. Future work will involve refining these definitions, expanding the methodology to other qualitative conversational aspects, and enhancing the fine-tuning process to enable simultaneous control over multiple features. The ultimate goal is to improve the reliability and adaptability of conversational LLMs, making interactions more transparent and precisely tailored to user needs. You can read the full research paper here.
