Dynamic Few-Shot Learning for Knowledge Graph Question Answering

The advent of large language models has opened up new avenues for Question Answering over Knowledge Graphs (KGQA), a field that has witnessed significant advancements due to their capabilities. Despite these advancements, large language models are not inherently designed for query generation, necessitating the development of fine-tuning solutions or ad-hoc architectures, which, while effective in certain scenarios, exhibit limitations in generalizing across diverse domains. In response to these challenges, this thesis presents a groundbreaking method known as Dynamic Few-Shot Learning (DFSL). DFSL leverages the principles of in-context learning combined with semantic similarity measures to create a versatile and robust framework for KGQA. An extensive evaluation across multiple benchmark datasets and architecture configurations, reveals not only the superior performance of DFSL in comparison to existing state-of-the-art methods but also exhibits adaptability against different KGs. This highlights the potential of DFSL to serve as a universally applicable solution for KGQA.