Knowledge Graphs (KGs) are structured representations of factual information, typically stored as (subject, predicate, object) triples. They are widely used to organize domain-specific knowledge and support fact-based reasoning.
Recent advances in Large Language Models (LLMs) have shown that combining them with external knowledge sources such as KGs can significantly improve performance in question answering (QA) tasks. This hybrid approach, known as Retrieval-Augmented Generation (RAG), helps reduce factual errors by grounding responses in reliable data.
However, while KGs are effective for answering factual questions, they struggle with more complex queries that require summarization, synthesis, or sensemaking. This limitation arises because KGs encode knowledge locally, making it difficult to extract broader, interconnected insights.
This project will explore how to improve context retrieval from KGs for different types of QA tasks. The focus will be on empirically answering the question: Given a Knowledge Graph and a text query, what is the most effective method for retrieving relevant knowledge to answer the query, depending on the type of question, e.g., factual, or summarisation?
To address this, this project will investigate retrieval strategies at varying levels of granularity such as individual entities, paths, and subgraphs, and evaluate their effectiveness across different QA scenarios.
School
Computer Science and Engineering
Research Area
Natural language processing (NLP) | Knowledge representation | Artificial intelligence
Suitable for recognition of Work Integrated Learning (industrial training)?
No
- Research environment
- Expected outcomes
- Supervisory team
- Reference material/links
The student will have access to infrastructure and computing resources to launch the project.
The supervisory team includes experts in NLP and Graph Machine Learning (GML).
Students will meet with their supervisor weekly to review progress.
In addition to gaining research experience in NLP, KG-RAG, model evaluation and benchmarking, there are two main deliverables:
- A comprehensive technical report describing different KG-RAG context retrieval approaches, their trade-offs especially across QA task types, and careful evaluation and analysis of results on suitable benchmark datasets; and
- Well designed and structured Python code implementing the retrieval and evaluation algorithms.
