CentralBank-AI-Classifier: Detecting AI vs. Non-AI Sentences in Central-Bank Discourse

CentralBank-AI-Classifier is a binary sentence-level classifier (AI, Non-AI) trained on BIS central-bank speeches. The model identifies whether a sentence is about AI (e.g., AI/ML/LLM/GenAI/NLP/vision topics) or not. It is built on the domain-adapted encoder CentralBank-BERT , which was pretrained on ~66M tokens from 2M+ sentences of BIS speeches (1996–2024).

Dataset

• Total labeled sentences: 3,245
• Class balance: AI = 1,619 | Non-AI = 1,626
• Grouping: URL-grouped to avoid speech leakage.
• Split (80/10/10 by URL): Train 2,603 · Dev 322 · Test 320

Labels were curated via rule-based retrieval (domain dictionary) followed by manual audit.

Training

• Base model: CentralBank-BERT
• Head: BertForSequenceClassification(num_labels=2)
• Max length: 128
• Optimizer: AdamW
• LR: 2e-5 · Weight decay: 0.01 · Warmup: 10%
• Batch: 16 (train) / 32 (eval)
• Epochs: up to 4 (early stopping on dev macro-F1)
• Precision: fp16 when available
• Threshold tuning: decision threshold selected on dev by macro-F1 sweep → τ = 0.05
• Loss: standard cross-entropy (dataset is balanced)

Evaluation (Held-out Test)

Metric	Value
Accuracy	0.9812
Macro-F1	0.9812
F1 (AI)	0.9810
F1 (Non-AI)	0.9815
ROC-AUC	0.9932
PR-AUC (AI as positive)	0.9959

Notes. Threshold τ = 0.05 was tuned on the dev set for macro-F1 and then fixed for test.

Reliability Check on Keyword-Retrieved Sentences

Scoring the keyword-retrieved corpus with the trained classifier (τ = 0.05):

• Mean P(AI) = 0.9877; Median = 0.9995
• Q1–Q3 = 0.9994–0.9995 (IQR = 0.0001)
• Predicted AI share = 0.9893
• High-confidence share (≥ 0.90) = 0.9872
• Borderline (±0.10 around τ, i.e., [0.00, 0.15]) = 0.0112

These statistics indicate the rule-based retrieval is highly reliable; only a small tail merits manual spot-checks.

Intended Use

• Filtering and measuring AI-related discourse in central-bank communications (speeches, testimonies, reports).
• Pre-filtering before downstream tasks (stance, sentiment, topic modeling).
• Corpus construction and time-series indicators of AI attention.

Out of scope: social media, consumer product reviews, or informal text.

---

Project GitHub Repository

The complete reproducible workflow, including the FinAI dictionary, dictionary-based tagging, AI sentence classification, sentiment analysis, and structural topic modeling, is available on GitHub:

CentralBank-AI: https://github.com/bilalezafar/CentralBank-AI

---

Usage

Simple pipeline

from transformers import pipeline

clf = pipeline("text-classification",
               model="bilalzafar/CentralBank-AI-Classifier",
               return_all_scores=False)

s = "We are piloting large language models to streamline supervisory analytics."
print(clf(s)[0])   # -> {'label': '1', 'score': 0.999...}
#Note Label_1=AI, Label_0=Non-AI

---

Citation

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}
}

---