Quantifying Legal Risk in Corporate Speech — Outcome-Linked Quotes Dataset

Quantifying Legal Risk in Corporate Speech — Outcome-Linked Quotes Dataset This dataset contains curated JSONL entries of corporate speech extracted from major U.S. litigation and regulatory proceedings, each linked to subsequent legal outcomes. Records include provenance fields (CourtListener docket/document IDs, source URLs, and jurisdiction metadata), short text quotations, labels, engineered features, and model predictions. Records: 24,462 quotes Cases: 125 cases Time span: 2000–2025 Jurisdictions: 10+ U.S. states Outcomes: Ranging from \$32k settlements to \$5B+ judgments File size: ~1 GB compressed, 3 GB Fields in each row: Embeddings: Large float array (CORAL embeddings). CORAL predictions: bucket, class, confidence, probabilities, scores. Features: 200+ interpretable and engineered features (lexical, sequential, linguistic, structural, legal risk). Metadata: case_id_clean (e.g., "1:19-cv-02184_dcd"), _metadata_src_path, docket_id, courtlistener_id, source_url, jurisdiction. Leakage prevention wrappers: separate quote, speaker, and context fields. Source: Derived from CourtListener (Free Law Project) via the RECAP Archive and API. Only metadata and short quotations necessary for analysis are included. Full filings are not redistributed; entries link back to CourtListener. Use case: Enables research on NLP methods for legal risk quantification, outcome prediction, and explainable AI in corporate compliance, litigation discovery, and insurance premium modeling. Citation: Jacob Dugan (2025). Quantifying Legal Risk in Corporate Speech — Outcome-Linked Quotes Dataset. Zenodo. https://doi.org/10.5281/zenodo.16934610 Attribution: CourtListener (Free Law Project). https://www.courtlistener.com Citation guide: https://www.courtlistener.com/cite/ License: CC BY 4.0 for annotations and derived labels. Original filings remain under CourtListener terms of service and applicable court rules.