Tús Maith: A Framework to Curate Natural Language Entry Points to Address the Initial Exploration Problem in RDF Knowledge Graphs
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Trinity College Dublin. School of Computer Science & Statistics. Discipline of Computer Science
Access
openAccess
Embargo end date
Citation
McNamara, Claire Louise, Tús Maith: A Framework to Curate Natural Language Entry Points to Address the Initial Exploration Problem in RDF Knowledge Graphs, Trinity College Dublin, School of Computer Science & Statistics, Computer Science, 2026
Abstract
Knowledge graphs (KGs) provide a powerful means of representing and linking rich complex information, yet their structure, ontologies, and query mechanisms create significant barriers for lay users attempting to begin exploration of them. This thesis investigates the challenge faced by lay users at the moment of first contact with an unfamiliar complex KG and introduces the Initial Exploration Problem (IEP), defined as a pre-goal state characterised by three interrelated barriers: scope uncertainty, ontology opacity, and query incapacity. These barriers prevent users from identification of a meaningful starting point for exploration and expose a limitation in many existing KG exploration interfaces which typically presuppose that the user already knows where to begin.
To address this problem, the thesis introduces Tús Maith, a multistage framework designed to support the creation of curated natural language question and answer (CuQA) entry points into KGs. These CuQAs function as cognitive and linguistic scaffolds that reveal both the scope and structure of a KG while abstracting away from ontology complexity and formal query languages. The framework is built around the lifecycle of the CuQA through three stages: Competency Question (CQ) template derivation from the KG schema, CQ template instantiation using KG entities, and domain expert curation. While CQs are traditionally used as ontology engineering artefacts, they are repurposed in this research as exploratory instruments capable of exposing the scope and structure of a KG through natural language.
Tús Maith is instantiated as TMv1, which incorporates large language models (LLMs) to support scalable generation and instantiation of CQ templates while foregrounding human oversight through a human-in-the-loop curation process. Two technical experiments investigate the capacity of LLMs to derive CQ templates from the KG schema and to instantiate these CQ templates under strict constraints of entity fidelity and hallucination control. Three user evaluations examine the effectiveness of CuQAs as exploration starting points for lay users and the viability of Tús Maith to support domain experts using the Virtual Record Treasury of Ireland (VRTI) KG. Results indicate that CuQAs are associated with positive post-use perceptions of scope revelation and approachability of the KG when compared to a neutral reference point, providing partial support for their ability to mitigate the IEP. Evaluation with domain experts further demonstrates that TMv1 can function as a viable system for curating such entry points. A proof-of-concept application to the LetterSampo correspondence KG provides supporting evidence of the portability of Tús Maith, as instantiated using LLMs becoming TMv1, beyond the VRTI KG across other digital humanities KGs.
This thesis makes one major and two minor contributions. The major contribution is the design, instantiation, implementation, and evaluation of Tús Maith, a framework that implements the lifecycle of CuQA entry points into KGs and demonstrates how such entry points can support early-stage exploration by lay users. The first minor contribution is the identification and characterisation of the Initial Exploration Problem (IEP), providing a conceptual framework that unifies three barriers (scope uncertainty, ontology opacity, and query incapacity) and offers a vocabulary for analysing first-contact KG exploration and designing interventions targeting this stage of interaction. The second minor contribution is an empirical investigation of LLM-assisted scaffolding for KGs, demonstrating how LLMs can support the scalable generation and instantiation of CQ templates while requiring domain expert oversight.
Together, these contributions advance methods for improving the accessibility of complex KGs and foreground the importance of supporting the earliest stage of user interaction: determining where exploration can begin.
Tús Maith Leath na hOibre - a good start is half the work.
Description
APPROVED
Endorsement
Review
Supplemented By
Referenced By
Sponsor: chool of Computer Science and Statistics Trinity College Dublin
Sponsor: Irish Research Council (IRC)
Author's Homepage: https://tcdlocalportal.tcd.ie/pls/EnterApex/f?p=800:71:0::::P71_USERNAME:MCNAMACL
Publisher: Trinity College Dublin. School of Computer Science & Statistics. Discipline of Computer Science
Type of material: Thesis

