Augmentation of large language model capabilities with knowledge graphs

Abstract

This postgraduate thesis explores the possibility for augmentation of the abilities of Large Language Models (LLMs) in the task of Question Answering by incorporating the technique of Retrieval-Augmented Generation (RAG) in conjunction with Knowledge Graph triples. Leveraging the structured knowledge representation of semantic triples, the study aims to enhance the contextual understanding and precision of LLMs in responding to user queries. The methodology proposed consists of a pipeline for the extraction of triples from source documents (Wikipedia articles) using a fine-tuned Language Model, their subsequent storage and retrieval through RAG, and their inclusion as contextual information to assist the LLM with providing grounded and reliable answers which are based on the provided sources. The findings suggest that while the answers show an improvement compared to unaided LLM generation, this specific methodology is not competitive with more conventional RAG applications.