CCF Database: A Machine-Learning-Annotated Corpus of 283,964 Canadian Climate Articles (1978–2026)

Abstract

News media shape how societies perceive and respond to climate change, yet fine-grained analysis remains hindered by fragmented corpora and ad hoc coding schemes. The Canadian Climate Framing (CCF) database answers both with a large-scale bilingual annotated corpus and a generalizable annotation framework. It covers 283,964 climate-related articles from 22 Canadian news outlets (1978–2026), processed into 9,908,776 two-sentence units (83.4% English, 16.6% French). Each unit carries 65 hierarchical binary annotations spanning eight main frames, actors, events, solutions, tone, geography, and urgency, produced by 128 BERT and CamemBERT classifiers trained on 4,000+ expert-coded sentences. Against a 1,000-sentence gold standard double-coded by an independent annotator (Cohen’s κ  = 0.596, Krippendorff’s α  = 0.698, Gwet’s AC1  = 0.894), the classifiers reach a pooled macro F 1 of 0.866 (mean per-category macro F 1  = 0.773), and every category carries an A/B/C reliability tier. The database—sentence-level annotations, per-article aggregations, and 10,192,740 BAAI/bge-m3 embeddings in six relational tables—is openly released on Zenodo (CC-BY 4.0) as a PostgreSQL edition and an Apache Parquet mirror, refreshed roughly every six months by the CCF Observatory.