Knowledge engineering for semantic understanding

This thesis addresses the challenges of improving semantic understanding in conversational agents by combining Knowledge Graphs (KGs) and Large Language Models (LLMs) within a flexible, multi-domain knowledge plugin architecture. We explore the inherent difficulties LLMs face in interpreting plain-text user queries, as well as the limitations of generative AI, particularly its tendency toward “hallucination” when generating responses. To mitigate this, we examine the complex process of extracting and structuring knowledge from raw text to construct KGs that serve as authoritative, context-rich foundations for information retrieval. The knowledge plugin architecture developed in this work enables conversational agents to leverage both KGs and LLMs to interact accurately with reliable, domain-specific sources. Our approach includes techniques such as fine-tuning and intelligent fewshot prompting to enhance LLMs’ ability to generate accurate, context-aware queries and responses over KGs. This integration significantly advances the potential for scalable, adaptable conversational agents capable of reliable information retrieval across multiple domains. The insights and techniques outlined in this thesis mark a critical step toward creating domain-agnostic AI systems that deliver semantically precise and trustworthy information.