M.S. Candidate: Barış Deniz Sağlam
Program: Data Informatics
Date: 29.08.2024 / 10:00
Place: A-212

Abstract: This thesis explores the use of small to medium-sized large language models (LLMs) for multi-hop question answering. As computational resources and latency present significant constraints in real-world applications, smaller language models (LLMs) are often utilized. However, these smaller models generally lack the extensive parametric knowledge and advanced reasoning capabilities possessed by their larger counterparts, such as GPT-4. This research investigates various augmentation strategies, notably the use of knowledge graphs, which provide a structured representation of facts and relationships to compensate for these limitations. This study investigates: whether knowledge graphs improve multi-hop question-answering capabilities of LLMs, the impact of integrating entity-relation triplets with textual content, and whether adaptation methods such as supervised fine-tuning or reinforcement learning with task-specific feedback improve the joint entity-relation extraction performance. The study introduces a novel prompting technique, Connect-the-Entities (CTE), which facilitates the extraction of relevant entity-relations before answering questions, thereby improving performance on the MuSiQue-Ans dataset with reduced computational demand. Additionally, the use of a pre-built knowledge graph as an external knowledge source demonstrates comparable results to baseline systems. Overall, this thesis contributes to the field by demonstrating how smaller LLMs can achieve enhanced question-answering performance through structured knowledge integration and advanced prompting techniques.