TACTICAL: A framework for building Wikipedia-derived timelines of atomic changes

Borkakoty, Hsuvas and Espinosa-Anke, Luis ORCID: https://orcid.org/0000-0001-6830-9176 2025. TACTICAL: A framework for building Wikipedia-derived timelines of atomic changes. Presented at: 28th European Conference on Artificial Intelligence, Bologna, Italy, 25-30 October 2025. Published in: Lynce, Inês, Murano, Nello, Vallati, Mauro, Villata, Serena, Chesani, Federico, Milano, Michela, Omicini, Andrea and Dastani, Mehdi eds. ECAI 2025. Frontiers in Artificial Intelligence and Applications IOS Press, pp. 4410-4417. 10.3233/faia251339

PDF - Published Version Available under License Creative Commons Attribution Non-commercial. Download (810kB)

Abstract

The well-known temporal misalignment in large language models (LLMs) emerges when they fail to recall temporal information. This is due to their training process, which happens without any explicit temporal grounding. To mitigate this issue, multiple approaches have been proposed, including fine-tuning on up-to-date data, retrieval augmented generation – where an LLM is directed to a recent dataset – or modifying an LLM’s knowledge via knowledge editing. Regardless of the method, however, the question of building datasets that accurately and faithfully reflect changes to events or entities remains open. Doing this in free text form and not only as triplets is desirable because LLMs benefit downstream from more context and can capture more nuanced relationships and cascading knowledge updates. Resources like Wikipedia can be leveraged for this thanks to their revision histories, which are expressed in free text and are both less biased and more comprehensive than knowledge graphs like Wikidata. In this paper, we propose TACTICAL, a methodology for creating timelines of Wikipedia entities and events, represented as revision pairs extracted from a wikititle’s timeline, and are categorized according to the atomicity of the changes affecting such entities or events. Our results suggest that LLMs struggle to recall event and entity timelines, even if they have seen them during pretraining. TACTICAL, on the other hand, proves to be an effective method for building temporally grounded datasets that are, in turn, effective tools for activating LLMs’ temporal knowledge.

Item Type:	Conference or Workshop Item (Paper)
Status:	Published
Schools:	Schools > Computer Science & Informatics
Publisher:	IOS Press
ISBN:	9781643686318
Date of First Compliant Deposit:	4 November 2025
Last Modified:	04 Nov 2025 11:45
URI:	https://orca.cardiff.ac.uk/id/eprint/182099

Actions (repository staff only)

Edit Item