Turning documents into data with dictionary methods and large language models, applied to real Federal Reserve, IMF, and World Bank text. Two labs, a responsible-use session, and a closing research workshop — all run in claude.ai.
The document pack introduced; document-term matrices; what LLMs change; the FOMC deliberation study.
Loughran–McDonald scoring of real Fed language; auditable word lists; the failure gallery (negation, hedging, domain shift); build-your-own dictionary.
Embeddings and attention without equations; the Fedspeak evidence (Hansen–Kazinnik; Trillion Dollar Words); live hawk/dove classification; what LLMs still get wrong.
Zero-shot vs. rubric prompts; 3-run stability test; benchmark vs. human labels; structured extraction with 'not stated' rules; adversarial framing.
Extraction schema with mandatory verification; the hedging ladder; two-institution comparison; a one-page policymaker brief.
Hallucination, reproducibility, bias, confidentiality; disclosure sentences; NIST / OECD / EU AI Act; research practice as internal control.
One-slide study template; cohort-matched project ideas; two generalized starter prompts; the take-home toolkit.
Morning 9:00–12:00 · Afternoon 1:00–5:00 · Two labs plus rules-of-the-game and a closing workshop. Head to the Labs page to run the prompts in Claude.