描绘央行行长思想的空间（英）

Mappingthe space ofcentral bankers’ ideas by Taejin Park, Fernando Perez-Cruz and Hyun SongShin Monetary and Economic Department October 2025 JEL classification: E52, E58, C55, C38Keywords: central bank communication, central bankspeeches, AI, topic modeling; embeddings BISWorking Papers are written by members of the Monetary and EconomicDepartment of the Bank for International Settlements, and from time to time by othereconomists, and are published by the Bank. The papers are on subjects of topicalinterest and are technical in character. The views expressed in this publication arethose of the authors and do not necessarily reflect the views of the BIS or its membercentral banks. This publication is available on the BIS website (www.bis.org). ©Bank for International Settlements 2025. All rights reserved. Brief excerpts may bereproduced or translated provided the source is stated. Mapping the Space of Central Bankers’ Ideas Taejin Park∗Fernando Perez-Cruz∗Hyun Song Shin∗ Abstract This paper explores the landscape of economic ideas as revealed in the machinelearning embedding of a comprehensive dataset of central bank speeches. Thisdataset, maintained by the BIS, encompasses 19,742 speeches delivered by almost1,000 officials from over 100 central banks over a period spanning three decades,from 1997 to 2025.As well as topic analysis of speeches at any moment intime, the evolution of the topics over time provides insights into how the focusof central bank thinking has been shaped by shifting policy challenges since1997.Parsing the embedding both through topics and through time providesrich insights into how economic ideas have taken shape through communicationpractices of central banks worldwide. To demonstrate its utility, we have conducteda series of analyses that map the global landscape of monetary policy discourse.Furthermore, we construct a quantitative framework—referred to as the "spaceofcentral bankers’ ideas"—which uncovers institutional patterns and highlights shiftsin policy approaches over time. 1Introduction Central bank communication is an integral part of central bank policy setting, influencing and aligningmarket behaviour with central bank assessments (Blinder et al. (2008)). Ben Bernanke, former chairof the Federal Reserve System once quipped that “monetary policy is 98 percent talk and 2 percentaction,” underscoring the outsized role of what central bankers say in guiding expectations (Bernanke(2015)).Researchers have taken this insight to heart, increasingly leveraging natural languageprocessing (NLP) to systematically analyse central bank communications. By converting speeches,meeting minutes, press releases and other texts into data, researchers can quantify tone, topics andsignals embedded in policymakers’ words (e.g., Hansen and McMahon (2016); Gorodnichenko et al.(2023)). The methods for analysing central bank communication have evolved alongside advances in NLP. Thecore approaches can be grouped intosentiment analysistechniques (measuring the tone or sentimentof communications) (Pang and Lee (2008)) andtopic modellingtechniques (identifying discussedthemes or topics) (Vayansky and Kumar (2020)). Often these methods are combined to create richertext-derived indicators. Early studies often relied on manual coding or simple word-count indices. A classic approach isdictionary-based sentiment analysis, using predefined word lists to measure tone. For instance,researchers frequently use the Loughran and McDonald (2011) financial lexicon to quantify positiveor negative sentiment in policy texts. While such dictionaries are transparent and easy to apply,they ignore context and nuanced language. Subsequent work introduced statistical topic modellingto uncover themes in central bank communications. Latent Dirichlet Allocation (LDA) (Blei et al.(2003)) became a standard tool and has been applied to central bank speeches to identify prevalenttopics. Hansen et al. (2018), for example, used topic modelling to study transparency in U.S. FederalReserve (Fed) communications. These models, however, treat text in a “bag-of-words” manner, lacking consideration of word order or context. They can miss subtle linguistic signals and their staticvocabulary makes it hard to capture new jargon or changing communication patterns. In recent years, NLP in economics has embraced more sophisticated machine learning techniques.Rather than relying solely on pre-defined dictionaries or bag-of-words counts, researchers startedharnessing predictive models and contextual embeddings that better capture the nuances of language.One milestone was the use of word embeddings. Early embedding methods like Word2Vec (Mikolovet al. (2013)) and GloVe (Pennington et al. (2014)) learned dense vector representations of words,allowing researchers to measure semantic similarities (e.g., “inflation” close to “prices” in vectorspace). Central bank researchers have trained such models on policy corpora to map the “language