This repository hosts the end-to-end workflow for analyzing AI discourse in central bank communications. It is organized into modular pipelines that map to the paper’s methods: dictionary construction → rule-based tagging & statistics → binary AI detector → three-class sentiment → structural topic modeling (STM).
| Folder | Description (files in folder + brief notes) |
|---|---|
| 1-FinAI Dictionary/ | Pipeline: FinAI Dictionary Construction Pipeline — builds the domain-specific AI dictionary from finance literature (merge Scopus exports, clean text, TF-IDF unigrams, bigrams/trigrams). Outputs the FinAI Dictionary (final, after manual review) for downstream use.. |
| 2-FinAI Dictionary Tagging and Statistics/ | Pipeline: FinAI Dictionary Tagging and Statistics Pipeline — rule-based speech & sentence tagging using the FinAI dictionary and production of descriptive statistics. Includes BIS data acquisition (with manual country annotation). Outputs: annual speech totals and AI shares; subdomain counts by year; sentence-level AI mentions by country; country-level speech totals; country × subdomain matrices; author/speaker AI-mention tallies; keyword/subdomain hit statistics. |
| 3-CentralBank-AI-Classifier/ | Pipeline: CentralBank-AI Classifier Pipeline — binary AI vs. non-AI sentence detector for coverage validation and high-precision filtering. Trains a Hugging Face Trainer on the base encoder bilalzafar/CentralBank-BERT with grouped splits by URL, class-weighted loss, threshold tuning, and full evaluation. Model on HF: bilalzafar/CentralBank-AI-Classifier. |
| 4-CentralBank-AI-Sentiment/ | Pipeline: CentralBank-AI Sentiment Pipeline — three-class sentiment (Positive / Neutral / Negative) for AI-tagged sentences. Data prep, class weights, early stopping, macro-F1 model selection, and evaluation tables. Uses base encoder bilalzafar/CentralBank-BERT. Model on HF: bilalzafar/CentralBank-AI-Sentiment. |
| 5-Topic Modeling/ | Pipeline: Structural Topic Modeling (STM) Pipeline — STM with country/year covariates, model-selection diagnostics (held-out likelihood, residuals, coherence/exclusivity), topic proportion plots, topic correlation matrices, and FREX word clouds. |
| LICENSE | MIT license for the repository. |
| Model | Purpose | Access link |
|---|---|---|
| CentralBank-AI-Classifier | Identify AI-related sentences in central-bank text. Fine-tuned on the bilalzafar/CentralBank-BERT backbone; performs binary classification (AI / non-AI) at the sentence level; intended to screen/filter large speech corpora. Input → Output: sentence → {AI / non-AI} + probability. | https://huggingface.co/bilalzafar/CentralBank-AI-Classifier |
| CentralBank-AI-Sentiment | Assess tone of AI-related sentences in central-bank text. Fine-tuned on the bilalzafar/CentralBank-BERT backbone; performs three-class sentiment labeling (Positive / Neutral / Negative) on sentences already identified as AI-related. Input → Output: AI sentence → {Pos / Neu / Neg} + probabilities. | https://huggingface.co/bilalzafar/CentralBank-AI-Sentiment |
Please cite as: Zafar, M. B., Ali, H., & Aysan, A. F. (2026). Signals from the Noise: Decoding Global AI Discourse in Central Bank Communications. Central Bank Review, Article 100268. https://doi.org/10.1016/j.cbrev.2026.100268
@article{zafar2026signals,
title = {Signals from the Noise: Decoding Global AI Discourse in Central Bank Communications},
author = {Zafar, Muhammad Bilal and Ali, Hassnian and Aysan, Ahmet Faruk},
year = {2026},
journal = {Central Bank Review},
pages = {100268},
doi = {10.1016/j.cbrev.2026.100268},
url = {https://doi.org/10.1016/j.cbrev.2026.100268}
}