laylaylo/global-leaders-discourses

WorldwideSpeechText

[Paper (placeholder)](placeholder) · [GitHub (placeholder)](placeholder) · [Dashboard](https://huggingface.co/spaces/laylaylo/global-leaders-discourses-dashboard)

Dataset Description

WorldwideSpeechText is a multilingual corpus of political speeches delivered by heads of state and government across 33 countries, covering 67 leaders, 47,750 unique speeches, and 71,677 total records including translations (1959–2026). The dataset is designed for longitudinal and comparative analysis of political rhetoric across regime types and world regions.

Each record is a single speech text sourced from official governmental archives and public records. Where a source publishes a speech in multiple languages, each language version is stored as a separate row sharing the same base speech_id (distinguished by a language suffix); the total record count therefore exceeds the unique speech count. Non-speech content — news articles, parliamentary question-and-answer sessions, debate transcripts, and interviews — was removed through source-specific heuristic filtering, retaining only prepared or delivered address texts.

More information on dataset construction can be found in the accompanying paper.

Countries & Leaders

Countries were selected to represent all four regime types in the V-Dem Institute's Regimes of the World (RoW) framework — Closed Autocracy, Electoral Autocracy, Electoral Democracy, and Liberal Democracy — and to provide at least one example per World Bank region where a qualifying country could be sourced. Within each country, priority was given to long-tenured leaders: extended time in office yields richer longitudinal material and tends to correlate with well-maintained official speech archives. Coverage was opportunistically extended to other leaders accessible through the same source when the marginal collection effort was small.

Per-country regime classifications, speech distributions, and interactive visualisations are available in the dataset dashboard.

Languages

Speeches appear in the original language of delivery. Several sources publish speeches in multiple languages (original and official translations); all available language versions are included. The dataset contains texts in 35 languages.

Language co-occurrence across the dataset is visualised in the paper's language chord diagram (paper/visuals/language_chord.png).

Dataset Structure

Loading

from datasets import load_dataset

# All countries combined
ds = load_dataset("laylaylo/global-leaders-discourses", "all", split="train")

# Single country
ds = load_dataset("laylaylo/global-leaders-discourses", "france", split="train")

Each country is a separate subset (config); the all config is their union. Every config has a single train split.

Fields

Dataset Creation

Source Data

Speeches were collected from official governmental websites, presidential office archives, and recognised public records. Collection used country-specific HTML parsers adapted to each source's structure. No automated crawling across the open web was performed; all sources are official or widely recognised archival repositories.

Processing

After extraction, non-speech records were removed using source- and country-specific heuristics — primarily filtering on document type labels, URL patterns, and structural features of the HTML. No translation, alignment, or further annotation was applied. Texts are preserved as published by the source.

Intended Use

This dataset is intended for:

• —Computational political science — studying how rhetoric varies across regime types, regions, and time
• —Multilingual NLP — large-vocabulary, long-form political text across diverse scripts and languages
• —Longitudinal analysis — tracking how individual leaders' language evolves over extended tenures

It is not intended as a complete archival record of any leader's speeches, nor as a source for real-time or journalistic use.

Considerations

Data Completeness

This dataset aggregates publicly available speeches from official archives. Coverage is necessarily incomplete: digital archives vary in quality, retrospective digitisation is uneven, and some speeches may not have been published online. We collected as comprehensively as each source permitted, but make no claim to exhaustiveness for any leader or date range, and do not represent this dataset as a new archive — it is a unified, computationally accessible aggregation of existing public records.

Copyright and Licensing Information

Speeches were collected from public governmental and official sources. The licensing terms for each source are listed below. Users are responsible for reviewing and complying with the terms of use of each source before using this dataset.

Disclaimer: We do not guarantee the accuracy of the legal information above and take no responsibility for any use that conflicts with applicable licences or laws. Users are solely responsible for ensuring their use complies with the relevant terms of service and applicable law in their jurisdiction.

Limitations

• —Speech counts vary substantially across leaders — from tens to tens of thousands — reflecting differences in source archive quality, a leader's total output, and tenure length rather than editorial selection within a collected archive.
• —Temporal coverage is densest from approximately 2015 onward, where governmental online archives are most comprehensive.
• —Some sources publish only selected speeches; completeness for a given leader depends entirely on what the source has made publicly available.

Citation

If you use this dataset, please cite:

@misc{WorldwideSpeechText,
  author    = {Yaayladere, Leyla and Jin, Zhijing},
  title     = {WorldwideSpeechText: A Multilingual Corpus of Political Leader Speeches Across Regime Types},
  year      = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/laylaylo/global-leaders-discourses}}
}

Acknowledgements

We would like to thank Aman Gokrani, Berelian Karimian, and Miu Takagi for their contributions to data collection.