Towards a unified user modeling language for engineering human centered AI systems

Technical Summary: Towards a Unified User Modeling Language for Engineering Human Centered AI Systems

Problem Statement
In the context of modern digital society and the rapid growth of Artificial Intelligence (AI), personalization has become essential for optimizing user experience and preventing digital inequalities. While AI enables sophisticated profiling—such as inferring fatigue via smartwatch data or predicting attitudes via Large Language Models (LLMs)—current engineering practices lack a standardized mechanism to represent user profiles. Existing solutions for user modeling are fragmented, often focusing on isolated aspects at a coarse level, and failing to provide a unified, machine-readable format. This fragmentation leads to incomplete profiles, interoperability issues, and a missed opportunity for fine-grained adaptation in model-driven engineering (MDE) and AI systems. There is currently no de facto standard user modeling language that integrates insights from the modeling community with perspectives from sociology and psychology.

Methodology
The authors propose a unified user modeling language designed to consolidate previous proposals and expand upon them. The development process involved:

1. Systematic Literature Review (SLR): A review of 30 papers on user modeling in MDE was conducted to identify existing abstract formalizations (metamodels or grammars). This analysis revealed a lack of unified standards and a tendency to favor static aspects over dynamic ones (e.g., mood, fatigue).
2. Metamodel Design: Based on the SLR and insights from other fields, the authors defined an abstract syntax formalized as a UML class diagram. This metamodel introduces a comprehensive set of user dimensions categorized into ten main groups: Relationship, Personal Information, Competence, Accessibility, Personality, Preference, Culture, Goals, Emotions, and Moods.
   • The model utilizes a custom Range datatype (0–100) to represent scores, scales (e.g., "bad_to_good"), and binary choices, offering flexibility for various dimension types.
   • Specific standards are integrated, such as the Common European Framework of Reference for Languages (CEFR) for language skills and Hofstede's cultural dimensions for culture.
3. Concrete Syntax Implementation: To ensure machine readability and compatibility with AI tools, the abstract syntax was translated into a JSON-based concrete syntax. A JSON Schema was derived from the metamodel to enforce data consistency and validity.
4. Proof of Concept (PoC): The language was implemented on the open-source BESSER low-code platform. A PoC was developed to demonstrate automatic personalization of a conversational agent. The system supports two approaches:
   • Direct Personalization: Embedding the user model (as a JSON object) directly into the LLM's context prompt.
   • Indirect Personalization: Using a secondary LLM to post-process pre-defined messages based on the user model.

Key Contributions

• Unified Metamodel: The paper presents a comprehensive metamodel covering 10 categories and numerous granular dimensions, surpassing previous works that typically covered a maximum of 5 categories and 14 dimensions. It explicitly includes dynamic aspects like mood and fatigue, as well as cultural and relational dimensions often omitted in prior models.
• Standardized JSON Syntax: By defining a JSON Schema derived from the metamodel, the authors provide a format that is both human-readable and natively understood by LLMs, facilitating the integration of user models into AI-driven pipelines without complex transformation steps.
• Tooling and PoC: The authors provide a working implementation on the BESSER platform, including a web-based form for creating user models and a personalized conversational agent interface. The PoC demonstrates the system's ability to adapt LLM responses (e.g., tailoring exercise recommendations) based on specific user constraints like age, disability, or language proficiency.

Results
The proof of concept successfully demonstrated the feasibility of the approach. In a specific scenario where an LLM was asked to generate muscle-gain exercise recommendations:

• For a 30-year-old male with paraplegia, the system automatically omitted lower-body exercises and focused on upper-body routines.
• For an 80-year-old male, the system prioritized light exercises, avoided weighted recommendations, and included safety advice to consult a doctor.
  These results indicate that the LLM could effectively interpret the structured JSON user model and adapt its output to align with the user's profile, thereby enhancing the relevance of the interaction.

Significance and Claims
The paper claims that this unified language addresses the critical gap in the MDE community by providing a complete, fine-grained, and interoperable representation of users. By moving beyond fragmented, static models, the proposed language enables a more exhaustive palette for representing users, which is a prerequisite for effective human-centered AI engineering. The authors argue that the demonstrated adaptation improves response quality and user experience, though they modestly note that a formal user study is required to empirically confirm these improvements.

The work lays the foundation for a pipeline where developers can model, test, and generate user-centered AI systems. Future work outlined by the authors includes interleaving this language with application modeling languages (e.g., IFML) to specify deterministic adaptations, implementing quality evaluation via model-based testing, exploring automatic profiling through implicit data, and ensuring strict adherence to ethical and privacy regulations like GDPR.