teagrjohnson/narrative-likert-roberta

Text Classification • Updated about 1 month ago • 24 • 1

Note NarraBert for Agency & Setting: A RoBERTa-base regression model that scores a passage on all 9 agency and setting dimensions (1–5) in a single forward pass. Distilled from Gemma labels and validated against human annotations. Pairs with the event-relation model to form NarraBert.

teagrjohnson/narrative-event-relation-roberta

Text Classification • Updated 3 days ago • 97 • 2

Note NarraBert for Event Relations: A RoBERTa-base classifier that labels adjacent event-trigger pairs for temporal ordering and causal relation, yielding passage-level temporal sequencing and causal density. Run downstream of the LitBank event detector (or whichever event detector model you prefer). See README for performance notes and pipeline prerequisites.

Note LLM-labeled annotations: 5K passages labeled by Gemma across the 11 dimensions, stratified by source and topic. Serves as the knowledge-distillation training set for NarraBert.

Note NarraDolma: ~3M passages across ~785K documents sampled from 12 Dolma sub-corpora, each with a 12-feature narrative vector (11 NarraBert dimensions plus event density). A stratified, narrative-enriched sample. Use for relative comparison across sources, not raw-Dolma prevalence.

Note Gold annotations: 400 passages human-annotated across all 11 narrative dimensions (agency, setting, event relations). The validation backbone for both the LLM labelers and NarraBert.

Note Arxiv paper