Abstract

Knowledge Graphs (KGs) provide structured, reliable, and domain-specific information across diverse fields, but accessing them requires formal query languages—a significant barrier for non-technical users. Large Language Models (LLMs) enable natural conversation, but are prone to hallucinations and lack grounding in up-to-date knowledge. This thesis bridges this gap by developing a universal framework that combines LLM-based natural language understanding with structured KG reasoning to enable conversational access to arbitrary KGs.
This thesis tackles this challenge through three progressive milestones. First, to establish the foundation for universal KG access, we eliminate expensive preprocessing on KGs and domain specific training. We develop KGQAN, the state-of-the-art KGQA system, which translates natural language questions into SPARQL queries for arbitrary KGs.
Second, evaluating conversational systems on KGs requires dialogue benchmarks based on KG information. We therefore develop CHATTY-GEN, which automates benchmark generation from arbitrary KGs through a multi-stage retrieval-augmented generation pipeline, reducing generation time by 99\%. Finally, to enable multi-turn conversational interaction with a KG, we develop CHATTY-KG, a modular multi-agent system that supports single-turn and multi-turn dialogues through specialized agents while maintaining low latency and high accuracy.
Extensive evaluations across diverse real-world KGs demonstrate that our approach significantly outperforms existing systems in answer quality, efficiency, and adaptability, establishing a comprehensive platform for natural, reliable, and scalable conversational access to KGs.