With better Wikipedia come better Large Language Models: A study on article reliability, temporal evolution, and community discussions

Borkakoty, Hsuvas 2025. With better Wikipedia come better Large Language Models: A study on article reliability, temporal evolution, and community discussions. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (PhD Thesis) - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (6MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (104kB)

Abstract

Wikipedia is the largest open multilingual knowledge-sharing platform, and a popular resource for both human readers and Natural Language Processing (NLP). However, its collaborative nature also makes it vulnerable. This thesis examines the challenges to reliability and content moderation in Wikipedia, with a focus on automated approaches that can support the community in maintaining high-quality content. The reader of this thesis will find, among others, experimental results on hoax articles, analyses of community discussions on Articles for Deletion (AfD), and novel frameworks based solely on Wikipedia that improve the temporal grounding of Large Language Modelss (LLMs). Our first contribution is Hoaxpedia, a dataset of Wikipedia hoax articles with textual contents and revision histories. We demonstrate that while text alone is an insufficient predictor for automatically identifying hoaxes, revision histories fill this gap. Hoax articles typically exhibit a markedly different activity timeline compared to legitimate articles. Then, we investigate Wikipedia’s community-based content moderation through the AfD process. We compile a multilingual and multi-platform dataset of deletion discussions and outcomes, and analyze prediction tasks for outcomes, stances, and policies. Additionally, we release a Python toolkit to support data collection and tasks related to outcome, stance, and policy prediction, based on our experiments. Finally, the third contribution of this thesis is the usage of Wikipedia passage pairs as demonstrations for prompting LLMs. We first propose Wikitide, an annotated dataset of definition pairs from two consecutive article snapshots. Using this dataset, we demonstrate that Language Model (LM) can be employed to distinguish between substantive information updates and surface-level updates (such as grammatical or paraphrasing) in article definitions. We then propose TACTICAL, a framework that automatically identifies meaningful changes in revision pairs of articles to probe and improve the temporal recall capabilities of LLMs.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Date of First Compliant Deposit:	24 July 2026
Date of Acceptance:	10 July 2026
Last Modified:	27 Jul 2026 13:10
URI:	https://orca.cardiff.ac.uk/id/eprint/188282

Actions (repository staff only)

Edit Item