CHEW: A Dataset of CHanging Events in Wikipedia

In this paper, we Since language models (LMs) are trained on raw deep dive on the notion of change by proposing web text and, often, without any explicit temporal CHEW (CHanging Events in Wikipedia), a temporally grounding (Zhao et al., 2024), they are prone to suffer grounded dataset from Wikipedia that focuses temporal misalignment (Luu et al., 2021; Lazaridou on finding important changes to events and entities, et al., 2021; Jang et al., 2022). While there is starting from a collection of Wikipedia events and a significant body of work concerned with fixing entities, and their associated changes over time extracted this issue via, e.g., in-domain pretraining (Gururangan from Wikipedia lists and originally curated et al., 2020), neologism-focused pretraining in the TAQA dataset (Zhao et al., 2024).