Language models encode extensive factual knowledge within their parameters. The accurate assessment of this knowledge is crucial for understanding and improving these models. In the literature, factual knowledge assessment often relies on cloze sentences, which can lead to erroneous conclusions due to the complexity of natural language (out-of-subject continuations, the existence of many correct answers and the several ways of expressing them). In this paper, we introduce a new interpretable knowledge assessment method that mitigates these issues by leveraging distractors—incorrect but plausible alternatives to the correct answer. We propose several strategies for retrieving distractors and determine the most effective one through experimentation. Our method is evaluated against existing approaches, demonstrating solid alignment with human judgment and stronger robustness to verbalization artifacts. The code and data to reproduce our experiments are available on GitHub.

La factualité des modèles de langue se dégrade avec le temps puisque les événements postérieurs à leur entraînement leur sont inconnus. Une façon de maintenir ces modèles à jour pourrait être la mise à jour factuelle à l’échelle de faits atomiques. Pour étudier cette tâche, nous présentons WikiFactDiff, un jeu de données qui représente les changements survenus entre deux dates sous la forme d’un ensemble de faits simples, sous format RDF, divisés en trois catégories : les faits à apprendre, les faits à conserver et les faits obsolètes. Ces faits sont verbalisés afin de permettre l’exécution des algorithmes de mise à jour et leur évaluation, qui est présentée dans ce document. Contrairement aux jeux de données existants, WikiFactDiff représente un cadre de mise à jour réaliste qui implique divers scénarios, notamment les remplacements de faits, leur archivage et l’insertion de nouvelles entités.

The factuality of large language model (LLMs) tends to decay over time since events posterior to their training are “unknown” to them. One way to keep models up-to-date could be factual update: the task of inserting, replacing, or removing certain simple (atomic) facts within the model. To study this task, we present WikiFactDiff, a dataset that describes the evolution of factual knowledge between two dates as a collection of simple facts divided into three categories: new, obsolete, and static. We describe several update scenarios arising from various combinations of these three types of basic update. The facts are represented by subject-relation-object triples; indeed, WikiFactDiff was constructed by comparing the state of the Wikidata knowledge base at 4 January 2021 and 27 February 2023. Those fact are accompanied by verbalization templates and cloze tests that enable running update algorithms and their evaluation metrics. Contrary to other datasets, such as zsRE and CounterFact, WikiFactDiff constitutes a realistic update setting that involves various update scenarios, including replacements, archival, and new entity insertions. We also present an evaluation of existing update algorithms on WikiFactDiff.