| --- |
| license: mit |
| base_model: |
| - bilalzafar/CentralBank-BERT |
| pipeline_tag: token-classification |
| tags: |
| - NER |
| - named-entity-recognition |
| - central-bank |
| - BIS |
| - speeches |
| - finance |
| - economics |
| - monetary policy |
| datasets: |
| - bilalzafar/BIS-Speeches-NER-dataset |
| language: |
| - en |
| metrics: |
| - f1 |
| - accuracy |
| library_name: transformers |
| --- |
| |
| # Central Bank-BERT for Named Entity Recognition (NER) |
| A **domain-adapted BERT model ([`CentralBank-BERT`](https://huggingface.co/bilalzafar/CentralBank-BERT))** was fine-tuned for **Named Entity Recognition (NER)** in central banking discourse. The model automatically identifies and labels key entities in central bank speeches and related documents, focusing on three categories of interest: |
| * **AUTHOR / SPEAKER** – the individual delivering the speech or statement |
| * **POSITION** – the official title or role of the speaker (e.g., Governor, Deputy Governor, Board Member) |
| * **AFFILIATION** – the institution or organization associated with the speaker (e.g., Bank of Japan, European Central Bank, Bank of England) |
|
|
| The **COUNTRY** label was not explicitly modeled, since this information can be reliably **inferred from the affiliation of the central bank**. |
|
|
| ## Data |
| * **Source**: **BIS database of central bank speeches (1996–2024)** |
| * **Corpus Size**: 17,648 speeches with 1,961 held out for validation. |
| * **Input Field**: *Speech descriptions*, which typically contain a short speech title along with the name, position, and institutional affiliation of the speaker. |
|
|
| **Annotation Process**: |
| 1. A subset of short speech descriptions was **manually annotated** with entity spans for Author, Position, and Affiliation. |
| 2. This annotated subset was used to **train an initial NER model**. |
| 3. The model was then applied to the larger dataset (1996–2024) to generate preliminary labels. |
| 4. All generated labels were **manually reviewed and corrected**, ensuring complete and consistent annotation across the entire corpus of available speeches. |
|
|
| This approach combined **manual expertise** with **machine-assisted annotation**, making it feasible to build a large-scale, high-quality dataset covering nearly three decades of central bank communication. |
|
|
| ## Data Preparation |
| 1. **Normalization**: Lowercasing, removal of diacritics, and unification of punctuation. |
| 2. **Alias resolution**: Institution abbreviations normalized (e.g., “BOJ” → “Bank of Japan”, “ECB” → “European Central Bank”). |
| 3. **Entity alignment**: Fuzzy string matching used to locate annotated entities in raw text. |
| 4. **BIO Encoding**: |
|
|
| * Tokenization with *BERT WordPiece tokenizer*. |
| * Conversion of annotations into **BIO tags** (`B-`, `I-`, `O`) at token level. |
| * Construction of a training file in **JSONL format** with `tokens` and `ner_tags`. |
|
|
| ## Model Training |
| * **Base model**: [`CentralBank-BERT`](https://huggingface.co/bilalzafar/CentralBank-BERT), a domain-adapted BERT trained on central banking corpora. |
| * **Task head**: Token classification layer with `num_labels = 7` (BIO scheme for Author, Position, Affiliation). |
| * **Token alignment**: Word-to-token mapping with subword label propagation (`-100` used for ignored positions). |
| * **Training setup**: |
| * Optimizer: AdamW with weight decay `0.01` |
| * Learning rate: `2e-5` |
| * Batch size: `16` (train & eval) |
| * Epochs: `3` |
| * Mixed precision (`fp16`) when available |
| * Evaluation with `seqeval` metrics (precision, recall, F1) |
|
|
| ## Results |
| The model was trained on **17,648 annotated speeches** with a **1,961-speech validation set**. Evaluation metrics are reported using **entity-level precision, recall, and F1-score** from the `seqeval` library. |
| **Final Validation Performance (Epoch 3):** |
| | Entity Type | Precision | Recall | F1-score | Support | |
| | --------------- | ---------- | ---------- | ---------- | ------- | |
| | **Affiliation** | 0.9850 | 0.9862 | 0.9856 | 1,734 | |
| | **Author** | 0.9816 | 0.9912 | 0.9864 | 1,936 | |
| | **Position** | 0.9735 | 0.9846 | 0.9790 | 1,942 | |
| | **Overall** | **0.9798** | **0.9862** | **0.9830** | — | |
| * **Accuracy (token-level):** 0.9956 | * **Overall F1 (macro):** 0.983 |
|
|
| The results show **high precision and recall across all three categories**, confirming that the model provides reliable structured metadata extraction from central bank communications. |
|
|
| --- |
|
|
| ## Other CBDC Models |
|
|
| This model is part of the **CentralBank-BERT / CBDC model family**, a suite of domain-adapted classifiers for analyzing central-bank communication. |
|
|
| | **Model** | **Purpose** | **Intended Use** | **Link** | |
| | ------------------------------- | ------------------------------------------------------------------- | ------------------------------------------------------------------- | ---------------------------------------------------------------------- | |
| | **bilalzafar/CentralBank-BERT** | Domain-adaptive masked LM trained on BIS speeches (1996–2024). | Base encoder for CBDC downstream tasks; fill-mask tasks. | [CentralBank-BERT](https://huggingface.co/bilalzafar/CentralBank-BERT) | |
| | **bilalzafar/CBDC-BERT** | Binary classifier: CBDC vs. Non-CBDC. | Flagging CBDC-related discourse in large corpora. | [CBDC-BERT](https://huggingface.co/bilalzafar/CBDC-BERT) | |
| | **bilalzafar/CBDC-Stance** | 3-class stance model (Pro, Wait-and-See, Anti). | Research on policy stances and discourse monitoring. | [CBDC-Stance](https://huggingface.co/bilalzafar/CBDC-Stance) | |
| | **bilalzafar/CBDC-Sentiment** | 3-class sentiment model (Positive, Neutral, Negative). | Tone analysis in central bank communications. | [CBDC-Sentiment](https://huggingface.co/bilalzafar/CBDC-Sentiment) | |
| | **bilalzafar/CBDC-Type** | Classifies Retail, Wholesale, General CBDC mentions. | Distinguishing policy focus (retail vs wholesale). | [CBDC-Type](https://huggingface.co/bilalzafar/CBDC-Type) | |
| | **bilalzafar/CBDC-Discourse** | 3-class discourse classifier (Feature, Process, Risk-Benefit). | Structured categorization of CBDC communications. | [CBDC-Discourse](https://huggingface.co/bilalzafar/CBDC-Discourse) | |
| | **bilalzafar/CentralBank-NER** | Named Entity Recognition (NER) model for central banking discourse. | Identifying institutions, persons, and policy entities in speeches. | [CentralBank-NER](https://huggingface.co/bilalzafar/CentralBank-NER) | |
|
|
|
|
| ## Repository and Replication Package |
|
|
| All **training pipelines, preprocessing scripts, evaluation notebooks, and result outputs** are available in the companion GitHub repository: |
|
|
| 🔗 **[https://github.com/bilalezafar/CentralBank-BERT](https://github.com/bilalezafar/CentralBank-BERT)** |
|
|
| --- |
|
|
| ## Usage |
| ```python |
| from transformers import pipeline |
| |
| # HF model repo |
| model = "bilalzafar/CentralBank-NER" |
| |
| ner = pipeline( |
| task="token-classification", |
| model=model, |
| tokenizer=model, |
| aggregation_strategy="simple" # merges subword pieces |
| ) |
| |
| # Example text |
| text = "Speech by Mr Yi Gang, Governor of the People's Bank of China, at the IMF Annual Meeting." |
| for ent in ner(text): |
| print(f"{ent['entity_group']:12} {ent['word']:<25} score={ent['score']:.3f}") |
| |
| # Example output: |
| # [{AUTHOR yi gang score=0.997}] |
| # [{POSITION governor score=0.999}] |
| # [{AFFILIATION people ' s bank of china score=0.999}] |
| ``` |
| --- |
|
|
| ## Citation |
|
|
| If you use this model, please cite as: |
|
|
| **Zafar, M. B. (2025). CentralBank-BERT: Machine learning evidence on central bank digital currency discourse. *Journal of Economics and Business.* [https://doi.org/10.1016/j.jeconbus.2026.106300](https://doi.org/10.1016/j.jeconbus.2026.106300)** |
|
|
| ```bibtex |
| @article{zafar2025centralbankbert, |
| title={CentralBank-BERT: Machine learning evidence on central bank digital currency discourse}, |
| author={Zafar, Muhammad Bilal}, |
| year={2026}, |
| journal={Journal of Economics and Business}, |
| url={https://doi.org/10.1016/j.jeconbus.2026.106300} |
| } |
| |
