Datasetpublic

bilalzafar/BIS-Speeches-NER-dataset

BIS Speeches NER dataset BIS Speeches NER dataset is a large-scale annotated dataset of central bank speeches (1996–2024) curated for Named Entity Recognition (NER). It was developed to support the fine-tuning of Central Bank BERT and related models for extracting structured metadata from central bank communication. Models Trained on This Dataset Central Bank NER – fine-tuned NER model for AUTHOR, POSITION, AFFILIATION. Dataset Schema Each row… See the full description on the dataset page: https://huggingface.co/datasets/bilalzafar/BIS-Speeches-NER-dataset.

sourceHugging Facemitupdated 1y agoView on Hugging Face

0likes37downloads

Dataset Card

BIS Speeches NER dataset

BIS Speeches NER dataset is a large-scale annotated dataset of central bank speeches (1996–2024) curated for Named Entity Recognition (NER). It was developed to support the fine-tuning of [`Central Bank BERT`](https://huggingface.co/bilalzafar/CentralBank-BERT) and related models for extracting structured metadata from central bank communication.

Models Trained on This Dataset

—Central Bank NER – fine-tuned NER model for AUTHOR, POSITION, AFFILIATION.

Dataset Schema

Each row corresponds to a single central bank speech record with the following fields:

Column	Type	Description
url	`string`	Direct BIS link to the PDF of the speech.
title	`string`	Short title of the speech as listed in the BIS archive.
description	`string`	Concise description of the speech, typically including the speaker’s name, role, and affiliation.
date	`datetime`	Date of the speech.
text	`string`	Full text of the speech.
processed\_text	`list[string]`	Pre-processed sentence-level text (tokenized, normalized, lowercased).
author	`string`	Name of the speech author or speaker.
affiliation	`string`	Institution or central bank associated with the speaker (e.g., People’s Bank of China, European Central Bank).
position	`string`	Official role or title of the speaker (e.g., Governor, Deputy Governor).
country	`string`	Country corresponding to the affiliation.

Source

—Original data: BIS database of central bank speeches (1996–2024).
—Corpus size: 17,648 speeches annotated for training, with 1,961 held out for validation.
—Language: English.

Annotation Process

1.A small subset of speech descriptions was manually annotated for entity spans (AUTHOR, POSITION, AFFILIATION).
2.This seed set was used to train an initial NER model.
3.The model was applied to the full BIS corpus (1996–2024) to generate preliminary labels.
4.All annotations were manually reviewed and corrected, ensuring full coverage and consistency.

This hybrid manual + machine-assisted annotation approach enabled efficient, large-scale dataset creation while maintaining high quality.

Example

json

{
  "tokens": ["speech", "by", "mr", "yi", "gang", ",", "governor", "of", "the", "people", "'s", "bank", "of", "china"],
  "ner_tags": ["O", "O", "B-AUTHOR", "I-AUTHOR", "I-AUTHOR", "O",
               "B-POSITION", "O", "O", "B-AFFILIATION", "I-AFFILIATION",
               "I-AFFILIATION", "I-AFFILIATION", "I-AFFILIATION"]
}

Intended Uses

—Training NER models for central banking and financial policy analysis.
—Benchmarking domain-specific NER tasks.
—Research on monetary policy communication and speaker metadata extraction.

Citation

If you use this dataset, please cite:

bibtex

@dataset{zafar2025centralbankner,
  author    = {Zafar, M.B.},
  title     = {BIS Speeches NER dataset},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/bilalzafar/BIS-Speeches-NER-dataset}
}