OntoChat: Reshaping Requirements Elicitation and Story Generation in Ontology Engineering with Large Language Models

This article introduces the OntoChat system—a tool that combines large language models with conversational interaction, designed to help ontology engineers efficiently generate high-quality user stories, reduce the cognitive burden of requirements elicitation, and enhance the participatory collaboration experience in ontology development. Addressing long-standing pain points in ontology engineering, the system optimizes processes through the concept of participatory requirements elicitation. Empirical validation confirms its advantages in user story quality and experience, providing directions for ontology engineering practice and research.

Ontology Engineering (OE) is a core field of knowledge representation, but it faces challenges in effectively capturing domain experts' knowledge. The traditional process has three major issues: 1. Domain experts lack ontological background, leading to communication barriers in requirements discussion; 2. Interviews/questionnaires produce scattered, unstructured information, making it time-consuming to convert into ontology elements and prone to losing details; 3. Static requirement documents struggle to adapt to the iterative and dynamic nature of ontology development.

OntoChat is based on large language models, with its core design concept being "participatory requirements elicitation": it uses conversational interaction to guide experts in gradually refining requirements, lowering the participation threshold; it integrates prompt engineering and domain knowledge to ensure generated content aligns with best practices in ontology modeling. The system architecture includes three main components: a dialogue management engine (maintains interaction state, identifies knowledge gaps, and follows up with questions), a user story generation module (converts dialogue information into standardized stories suitable for ontology engineering, ensuring consistency with existing ontologies), and a knowledge verification and feedback loop (presents stories for users to confirm and modify).

The team conducted a formative study with 24 participants (including ontology engineers and domain experts), comparing traditional methods with the OntoChat-assisted approach. Results show: User stories generated by the OntoChat group are superior in completeness, consistency, testability, and traceability, and they identified implicit requirements missed by traditional methods; user experience feedback is positive—domain experts recognize the naturalness of conversational interaction, and ontology engineers reduce the workload of requirements conversion.

The contributions of OntoChat include: 1. Promoting a new paradigm of participatory ontology engineering, enabling domain experts to deeply engage in the requirements elicitation phase; 2. Verifying the feasibility of LLM-assisted requirements engineering, proving it can act as an intelligent intermediary to facilitate cross-background communication; 3. Providing empirical insights into LLM prompt engineering strategies in specific domains, offering references for related research.

Practical implications: Introduce conversational AI to assist requirements elicitation, drawing on the concepts of progressive dialogue and real-time feedback; emphasize the role of user stories in ontology engineering; build prompt engineering capabilities. Future directions: Multilingual support, multimodal interaction, automatic ontology construction, collaboration functions, etc.