Which Text Method to Choose for Analysing Central Bank Communication? A Comparison of Artificial Intelligence and Previous Techniques

Our study compares the characteristics of text analysis methods on communications text samples of the US Federal Reserve and four Central and Eastern European central banks. Based on our results, methods based on BERT-type models are the most accurate at capturing the monetary policy, real economic and inflation information contained in central bank texts, outperforming OpenAI GPT-4.1 and GPT-5 models. BERT-type methods are faster than GPT models and can be run offline without a subscription, but their disadvantages are that the models require separate training, which entails hardware and labour costs, and that modifying the method is cumbersome. Conversely, GPT models are more flexible and have proven to be more accurate on new central bank samples. The dictionary-based methods used as benchmarks are significantly less accurate, but their use may be justified in certain cases due to their speed, cost-free operation and the transparency of the method.