Year: 2025
Type: dissertation
Abstract: This thesis investigates how different computational approaches can support relational search tasks, with a focus on biographical connections between historical individuals and places. The study compares three methods: structured retrieval from knowledge graphs (KGs), open-ended generation by large language models (LLMs), and retrieval-augmented generation (RAG), where LLMs are guided by structured context from KGs. The evaluation is situated in the domain of Finnish cultural history, focusing on artists of the Golden Age and their ties to cities such as Paris. Relational queries were manually crafted and tested across the three methods using a consistent prompt structure. Results show that KGs provide high precision and verifiability but are limited by data completeness. LLMs offer broad coverage and fluent responses yet frequently hallucinate connections or omit key facts—especially for lesser-known individuals. RAG improves output structure and accuracy but still requires careful prompt engineering and manual validation. The findings suggest that no single method is sufficient in isolation. Instead, KGs and LLMs are complementary: symbolic approaches excel in precision, while neural models enhance coverage and expressiveness. A small-scale demo application developed alongside the thesis demonstrates how these technologies can be combined to support semi-automated relational search. This work contributes to ongoing research in hybrid AI, cultural heritage data enrichment, and interpretable semantic retrieval.
Source: Aaltodoc (Aalto University)
Author: Malikova, Viktoria
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7076973871