Masterarbeit
Comparing Concepts for Producing Structured Outputs in an LLM Chatbot to Support International Student Onboarding
Completion
2025/10
Research Area
Students
Thi Ngoc Yen Nguyen
Advisers
Verena Traubinger M.Sc.
Dr.-Ing. Sebastian Heil
Description
Large Language Models (LLMs) have demonstrated strong capabilities in generating natural language responses; however, their outputs often consist of long texts without visual elements that may overwhelm users or hinder rapid comprehension. In contrast, traditional non-generative chatbots often rely on structured, interface-friendly elements such as buttons, cards, or carousels to present information succinctly and enable quick user interactions. These visual formats not only enhance clarity but also guide users toward specific actions, improving the overall user experience. Despite their linguistic strengths, LLMs have yet to generate structured outputs suitable for modern conversational interfaces consistently. This thesis investigates whether LLMs can produce such organized formats and whether these outputs meaningfully improve the clarity, efficiency, and quality of user interactions in real-world applications.
This experiment is using university onboarding as a case study. The approach involves exploring different strategies for structuring responses, such as defining output schema, using format-specific instructions, or fine-tuning prompt templates, to determine which methods produce consistent, interpretable results. Various implementation options will be reviewed and tested to identify the most effective way to guide LLMs toward generating structured formats with established UI elements. These outputs will then be tested in prototypical chatbots designed to support university onboarding, where users often seek clear, actionable information. The effectiveness, clarity, and user satisfaction of these different examples of structured responses in a real-world academic setting will be assessed with user tests.
The objective of this thesis is the creation of a solution or the combination of existing approaches to solve the problem of generating structured output from an LLM chatbot. This comprises the analysis of the state of the art of current LLM technologies, schemas and protocols, graphical chatbot interfaces, and other relevant literature. Prototypes for one use case should be implemented to test the different conceptualized approaches. A suitable evaluation has to be conducted which regards the feasibility and consistency of the different approaches on a technical level, and tests how they are evaluated by users, as well as their compliance with requirements which were extracted through the literature research.
