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

The article describes a project to build an NLP pipeline that can classify central bank communications into granular policy signals, such as hawkish or dovish language on inflation, interest rates, and economic activity.

Details

The project aimed to automatically track which central banks are turning hawkish or dovish based on their communications. The approach involves breaking central bank documents like policy statements, meeting minutes, and press conferences into individual sentences and classifying each one for sentiment (e.g., rate hike, rate cut, hawkish guidance) and topic (e.g., inflation, interest rates, labor market). This granular classification provides insights beyond just overall hawkish/dovish stances, such as a central bank having hawkish language on inflation but dovish language on the labor market. The pipeline has four stages: 1) Crawling central bank documents, 2) Splitting the text into sentences, 3) Classifying each sentence using bank-specific prompts and rules, and 4) Aggregating the sentence-level classifications into higher-level metrics. The author found that central bank language is domain-specific enough that generic sentiment analysis often fails, requiring custom prompts and rules to handle boilerplate language and counterintuitive dissent direction.