Building an NLP Pipeline to Classify 225,000 Central Bank Sentences

The Problem Central banks communicate through dense, jargon-heavy documents — policy statements, meeting minutes, press conferences. A single Fed statement is 1,500+ words. The ECB publishes minutes in 10,000+ word documents. Multiply that by 26 central banks, each publishing monthly or quarterly, and you have an impossible amount of text to track manually. I wanted to answer a simple question: which central banks are turning hawkish and which are turning dovish — right now? The Approach Instead of summarizing entire documents, I break them into individual sentences and classify each one. Every sentence gets two labels: Sentiment (what policy direction does it signal?): rate_hike , rate_cut , rate_hold guidance_hawkish , guidance_dovish dissent_hawkish , dissent_dovish liquidity_easing , liquidity_tightening neutral , irrelevant Topic (what economic area?): mp_inflation , mp_interest_rate , mp_economic_activity mp_labor_market , mp_exchange_rate , mp_credit financial_stability , fiscal