Vulnerable Identities Recognition Corpus (VIRC) for Hate Speech Analysis

Welcome to the Vulnerable Identities Recognition Corpus (VIRC), a dataset created to enhance hate speech analysis in Italian and Spanish news headlines. VIRC provides annotated headlines aimed at identifying vulnerable identities, dangerous discourse, derogatory mentions, and entities. This corpus contributes to developing more sophisticated hate speech detection tools and policies for creating a safer online environment. The work has been published at the Tenth Italian Conference on Computational Linguistics (CLiC-it 2024) with the name of Vulnerable Identities Recognition Corpus (VIRC) for Hate Speech Analysis. The code for the experiments performed in the paper are available in the github repo https://github.com/oeg-upm/virc.

Overview

VIRC is designed to support the study of hate speech in headlines from two languages: Italian and Spanish. It includes 880 headlines (532 Italian and 348 Spanish), collected and annotated with the following labels:

• Named Entities: Identifies persons, locations, organizations, groups, etc. mentioned in the headline.
• Vulnerable Identity Mentions: Labels groups such as women, LGBTQI, ethnic minorities, and migrants targeted by hate speech.
• Derogatory Mentions: Marks phrases that are derogatory towards vulnerable groups.
• Dangerous Speech: Highlights parts of the text perceived as potentially inciting hate or perpetuating harmful stereotypes.

Data

• Spanish: The Spanish datasets are split into two sets, agreement and disagreement. Agreement set contains the data annotated by the two original annotators, while the disagreement set contains the news where no agreement was reached and a third annotator was needed.
• Italian: The Italian data consists of only one set annotated by two annotators.

Annotation

The VIRC_Guidelines.pdf contains the annotation guidelines provided to annotators. This can be seen sintetized in the paper. The dataset is provided with several splits depending of which elements are included:

• Annotations (Spanish): Annotators annotations for Spanish.
• Annotations (Italian): Annotators annotations for Italian.
• Gold (Spanish): Gold standard annotations for Spanish.
• Gold (Italian): Gold standard annotations for Italian.
• Comments (Spanish): Annotators comments for Spanish.
• Comments (Italian): Annotators comments for Italian.

The different splits include:

Configuration	Annotations (Spanish)	Annotations (Italian)	Gold (Spanish)	Gold (Italian)	Comments (Spanish)	Comments (Italian)	Number of Rows
all	✅	✅	✅	✅	✅	✅	6027
all_tags	✅	✅	✅	✅	❌	❌	5536
spa_all	✅	❌	✅	❌	✅	❌	3407
ita_all	❌	✅	❌	✅	❌	✅	2620
spa_all_tags	✅	❌	✅	❌	❌	❌	3256
ita_all_tags	❌	✅	❌	✅	❌	❌	2280
gold_tags	❌	❌	✅	✅	❌	❌	1131
spa_gold_tags	❌	❌	✅	❌	❌	❌	550
ita_gold_tags	❌	❌	❌	✅	❌	❌	581
tags	✅	✅	❌	❌	❌	❌	4405
spa_tags	✅	❌	❌	❌	❌	❌	2125
ita_tags	❌	✅	❌	❌	❌	❌	2280
comments	❌	❌	❌	❌	✅	✅	491
spa_comments	❌	❌	❌	❌	✅	❌	151
ita_comments	❌	❌	❌	❌	❌	✅	340

Usage and Information

The dataset can be loaded with:

from datasets import load_dataset
dataset = load_dataset("oeg/virc")

BibTeX Entry and Citation Info

@inproceedings{guillen-pacho-etal-2024-vulnerable,
    title = "The Vulnerable Identities Recognition Corpus ({VIRC}) for Hate Speech Analysis",
    author = "Guill{\'e}n-Pacho, Ibai  and
      Longo, Arianna  and
      Stranisci, Marco Antonio  and
      Patti, Viviana  and
      Badenes-Olmedo, Carlos",
    editor = "Dell'Orletta, Felice  and
      Lenci, Alessandro  and
      Montemagni, Simonetta  and
      Sprugnoli, Rachele",
    booktitle = "Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC-it 2024)",
    month = dec,
    year = "2024",
    address = "Pisa, Italy",
    publisher = "CEUR Workshop Proceedings",
    url = "https://aclanthology.org/2024.clicit-1.50/",
    pages = "417--424",
    ISBN = "979-12-210-7060-6",
    abstract = "This paper presents the Vulnerable Identities Recognition Corpus (VIRC), a novel resource designed to enhance hate speech analysis in Italian and Spanish news headlines. VIRC comprises 921 headlines, manually annotated for vulnerable identities, dangerous discourse, derogatory expressions, and entities. Our experiments reveal that large language models (LLMs) struggle significantly with the fine-grained identification of these elements, underscoring the complexity of detecting hate speech. VIRC stands out as the first resource of its kind in these languages, offering a richer annotation schema compared to existing corpora. The insights derived from VIRC can inform the development of sophisticated detection tools and the creation of policies and regulations to combat hate speech on social media, promoting a safer online environment. Future work will focus on expanding the corpus and refining annotation guidelines to further enhance its comprehensiveness and reliability."
}

Acknowledgements

This work is supported by the Predoctoral Grant (PIPF-2022/COM-25947) of the Consejería de Educación, Ciencia y Universidades de la Comunidad de Madrid, Spain. Arianna Longo's work has been supported by aequa-tech. The authors gratefully acknowledge the Universidad Politécnica de Madrid (www.upm.es) for providing computing resources on the IPTC-AI innovation Space AI Supercomputing Cluster.

License

This work is licensed under the MIT License. For more details, see the LICENSE file.