Characterizing Narrative Content in Web-Scale Pretraining Data

This is the landing page for resources related to the "Characterizing Narrative Content in Web-Scale Pretraining Data" project. is a pair of RoBERTa-based models fine-tuned to detect narrative content in web text. The models span agency, setting, and event features. is a sample of the Dolma pretraining corpus with NarraBERT labels applied.

These resources can be used across many narratology applications. Because we operationalize narrative as a multi-dimensional, continuous construct, the models and datasets can be used to study narrative content at a fine-grained level across a wide variety of text, including fiction, non-fiction, and other narrative-adjacent genres. Contact Teagan Johnson with any questions or comments!