Bank for International Settlements publishes research using AI embeddings to map the evolution of central bank communication

The Bank for International Settlements published a working paper that builds a quantitative “space of central bankers’ ideas” by applying modern natural language processing to a large cross-country corpus of central bank speeches, aiming to map how monetary policy discourse clusters and evolves over time. The paper uses a BIS-maintained dataset of 19,742 English speech transcripts delivered by almost 1,000 officials from over 100 central banks between 1997 and March 2025, enriched with metadata such as speaker role and location. Speeches are standardised into bullet-point summaries using GPT-4o and converted into 3,072-dimensional embeddings (OpenAI text-embedding-3-large), which are then analysed with BERTopic (UMAP dimensionality reduction and HDBSCAN clustering) and visualised in a two-dimensional semantic map. The results highlight strong regional clustering, multiple distinct clusters for large central banks such as the Federal Reserve System and the European Central Bank (ECB), and peripheral clusters for newer themes including climate change and benchmark reforms, alongside time-patterns such as the persistent presence of inflation, a pronounced rise in climate-related communications in recent years, and more temporary prominence for Basel II implementation. As an institutional case study, separate topic modelling of ECB speeches groups communications into broad regions (monetary policy, financial stability, and euro area integration) and shows monetary policy communications clustering by policy regime over time; a complementary GPT-4o-based sentiment classification (hawkish/neutral/dovish, averaged across repeated runs) is used to validate that these temporal clusters align with shifts in policy stance. The authors suggest the framework could be extended toward real-time monitoring of communication shifts and combined with econometric analysis.