Persona-L has Entered the Chat: Leveraging LLMs and Ability-based Framework for Personas of People with Complex Needs

Research Background and Issues

• Identified Problems or Challenges: Current methods for creating user personas, especially for individuals with complex needs, tend to oversimplify or perpetuate stereotypes. This issue is particularly pronounced when addressing users with multiple disabilities (e.g., individuals with Down syndrome), as traditional persona creation methods struggle to comprehensively capture their dynamic needs and real-life contexts.
• Significance: Generating accurate and empathetic personas for individuals with complex needs can significantly enhance the empathy and precision of user experience design (UX Design), better addressing these users' requirements. Incorrect or incomplete persona descriptions may lead to ineffective or misleading design solutions.
• Research Motivation and Related Work:
  • Persona creation methods have evolved from qualitative approaches (e.g., interviews and observations) to data-driven methods (e.g., leveraging social media data), many of which rely on high-quality data.
  • Recent advancements in large language models (LLMs) have demonstrated strong performance in text generation tasks, but challenges such as stereotyping and overgeneralization persist when generating user personas.
  • Related research suggests that the Ability-Based Design Framework can more effectively depict users' capabilities and limitations, reducing biases associated with "disability."

Proposed Solution

• Solution: The Persona-L system, an interactive persona creation tool tailored for users with complex needs, developed by integrating LLMs (GPT-4o mini) with the Ability-Based Design Framework.
• Innovations:
  1. Ability-Based Persona Framework: Shifts the focus from solely emphasizing user limitations to highlighting their capabilities and the factors that enable or hinder these abilities in specific contexts.
  2. Interactivity: Users can not only generate personas but also interact with them through a chat interface, dynamically asking questions and receiving responses generated by LLMs and data sources.
  3. RAG Framework (Retrieval-Augmented Generation): Combines curated data sources (e.g., life experiences shared by Down syndrome communities) with LLM generation capabilities, linking responses to real-world data for improved accuracy.
• Implementation Steps and Key Technologies:
  1. Data Collection and Processing: Collect 80 real-life stories from public data sources (e.g., forums, organizations related to Down syndrome) and extract key persona attributes (e.g., age, occupation, ability drivers, limitations).
  2. User Interface (UI) Design: The system consists of three main phases:
     • Profile Phase: Users select demographic information, themes (e.g., employment, education, family), and personalized persona abilities.
     • Ability Phase: Displays ability drivers and limitations related to the selected theme, enriched by real-life stories for context.
     • Interaction Phase: Users engage in conversations with virtual personas to gain further insights through dialogue.
  3. LLM and RAG Integration: Utilize the retrieval-augmented generation framework, dynamically generating data-based responses by querying an embedded vector database (Chroma).

Research Outcomes

• Specific Results:
  1. The Persona-L system successfully generated personas relevant to individuals with complex needs, such as those with Down syndrome, across various themes and scenarios.
  2. Provided a dynamic chat interface that enhanced users' empathy and understanding of the personas.
  3. Validated the effectiveness of Persona-L through usability testing with six UX design experts, collecting feedback on persona transparency and language style improvements.
• Advantages Compared to Existing Methods:
  • Offers multidimensional user representation, balancing abilities and limitations to reduce stereotypes.
  • Enables dynamic persona generation and interaction, moving beyond traditional static document-style personas, enhancing realism and applicability.
  • Utilizes the RAG framework to mitigate data biases and hallucination issues inherent to LLMs.
• Experimental or Evaluation Results:
  • User experience studies indicated that participants found Persona-L improved their understanding of users with complex needs during research and design phases (e.g., understanding "ability barriers").
  • The system outperformed direct LLM queries in consistency, contextual relevance, and accuracy of generated dialogues.
  • Participants suggested further optimization of persona responses in terms of language complexity and length to better align with user expectations.
• Limitations and Future Directions:
  1. Data Transparency and Credibility: Current data validation relies heavily on manual curation; future work could explore integrating more real-time dynamic data updates.
  2. Long-Term Usage Behavior: Testing primarily focused on short-term interactions; future research should investigate how users utilize Persona-L in long-term design practices.
  3. Diversity and Customization: Persona diversity is currently limited by the breadth of the dataset; expanding data sources and enabling user co-creation and multi-persona dialogues could enhance persona representation.
  4. Addressing Potential Bias: Frameworks such as CoMPosT should be employed to further evaluate and mitigate biases in LLM-generated content.

Through Persona-L, this study demonstrates the potential of leveraging LLMs and ability-based frameworks to simulate users with complex needs, offering a more empathetic and practical tool for design. However, future work must delve deeper into data validity, persona diversity, and dynamic learning improvements.