EcuHate-es — Ecuadorian Spanish Hate Speech Dataset
The first labeled dataset for hate speech classification in Spanish produced in Ecuador.
Texts were collected from X (formerly Twitter), geolocated to Ecuadorian users across two regions
(Costa and Sierra), and annotated through a rigorous collaborative human annotation process.
This dataset is the result of research project FCI-036-2023, conducted at the Universidad de Guayaquil
(Faculty of Mathematical and Physical Sciences, Software Engineering Program, Ecuador)
Dataset details
Language
Spanish (es) — Ecuadorian variety
Source
X (Twitter), geolocated to Ecuador
Regions
Costa · Sierra
Task
Binary hate speech classification
Labels
YES (hate speech) · NO (no hate speech)
Size
7,348 texts
License
CC BY 4.0
Role
Name
Affiliation
Principal Investigator & Project Director (FCI-036-2023)
César Espin-Riofrio
Universidad de Guayaquil
Research Collaborator
Ignacio Valencia-Moyano
Universidad de Guayaquil
Research Collaborator
Allan Méndez-Vega
Universidad de Guayaquil
Data structure
Field
Type
Description
id
string
Unique text identifier
text
string
Original Spanish text from X
region
string
Ecuadorian region: Costa or Sierra
date
string
Tweet timestamp (ISO 8601)
num_tokens
integer
Token count of the text
label
string
YES = hate speech · NO = no hate speech
confidence
string
Annotation confidence: high or low
{"id":"d4e2e2f4","text":"Texto en español aquí","region":"Costa","date":"2023-01-26T13:34:25.000Z","num_tokens":34,"label":"YES","confidence":"high"}
Label distribution
Label
Meaning
Count
%
YES
Hate speech
4,114
56.0%
NO
No hate speech
3,234
44.0%
Total
7,348
Confidence distribution
Confidence
Criterion
Count
%
high
Both Doccano rounds agree (label2 == label3), including unanimous cases
4,580
62.3%
low
Doccano rounds disagree; original reference label acted as tiebreaker
2,768
37.7%
The confidence field reflects the degree of agreement between the two independent Doccano
annotation rounds. When both rounds agree, the final label is considered high confidence.
When they disagree, the original reference label (Gómez & Tapia, 2024) acted as tiebreaker
and the record is marked low confidence. Researchers may choose to use only high confidence
records (4,580 texts) for more conservative experiments.
Regional distribution
Region
Count
Costa
3,623
Sierra
3,725
Annotation
The raw corpus was compiled by Gómez & Tapia (2024) under research project FCI-036-2023 at
Universidad de Guayaquil — approximately 15,000 posts from X geolocated to Ecuadorian users,
reduced to 11,657 after deduplication.
Parameter
Value
Platform
Doccano (deployed on Microsoft Azure)
Annotators
200 (Law, Sociology, Political Science, Psychology — 8th–9th semester)
Krippendorff's Alpha, Fleiss' Kappa, Cohen's Kappa — consistent with hate speech annotation literature
Acknowledgements
The research team gratefully acknowledges the institutional support that made the annotation
process possible. We thank the deans and program directors of the Faculty of Jurisprudence,
Social and Political Sciences (Law, Sociology, and Political Science programs) and the
Faculty of Psychological Sciences (Psychology program) at Universidad de Guayaquil for
facilitating the participation of their students as annotators. We also thank the dean of the
Faculty of Mathematical and Physical Sciences for supporting the development of this project.
Citation
@dataset{espinriofrio2024ecuhate,
author = {Espin-Riofrio, César and Valencia-Moyano, Ignacio and Méndez-Vega, Allan},
title = {EcuHate-es: Ecuadorian Spanish Hate Speech Dataset},
year = {2024},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/cespinr/EcuHate-es}
}
