Year: 2025
Type: dissertation
Abstract: This thesis introduces two RAG pipelines that use KGs that are constructed frombiometric metadata, which is metadata associated with individuals. Such metadatais particularly valuable in contexts where the identity of the information sourceis critical, transcribed speeches or debates. The proposed KG-RAG pipelines areevaluated against a baseline RAG model based on cosine similarity. Experimentalresults demonstrate that the KG-based approaches consistently achieve superiorranking scores. Furthermore, evaluation using LLM-as-a-judge further supportsthe improved performance of the KG-RAG pipelines.
Author: Mattson, Marie
Institution: Uppsala University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7064588206