Skip to main navigation Skip to search Skip to main content

Towards Safer Hebrew Communication: A Dataset for Offensive Language Detoxification

  • Natalia Vanetik
  • , Lior Liberov
  • , Marina Litvak
  • , Chaya Liebeskind

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Text detoxification is the task of transforming offensive or toxic content into a non-offensive form while preserving the original meaning. Despite increasing research interest in detoxification across various languages, no resources or benchmarks exist for Hebrew, a Semitic language with unique morphological, syntactic, and cultural characteristics. This paper introduces HeDetox, the first annotated dataset for text detoxification in Hebrew. HeDetox contains 600 sentence pairs, each consisting of an offensive source text and a non-offensive text rewritten with LLM and human intervention. We present a detailed dataset analysis and evaluation showing that the dataset benefits offensive language detection. HeDetox offers a foundational resource for Hebrew natural language processing, advancing research in offensive language mitigation and controllable text generation.

Original languageEnglish
Title of host publicationProceedings of the 15th International Conference on Recent Advances in Natural Language Processing - Natural Language Processing in the Generative AI Era, RANLP 2025
EditorsGalia Angelova, Maria Kunilovskaya, Marie Escribe, Ruslan Mitkov
PublisherIncoma Ltd
Pages1289-1298
Number of pages10
ISBN (Electronic)9789544520984
DOIs
StatePublished - 1 Jan 2025
Externally publishedYes
Event15th International Conference on Recent Advances in Natural Language Processing - Natural Language Processing in the Generative AI Era, RANLP 2025 - Varna, Bulgaria
Duration: 8 Sep 202510 Sep 2025

Publication series

NameInternational Conference Recent Advances in Natural Language Processing, RANLP
ISSN (Print)1313-8502

Conference

Conference15th International Conference on Recent Advances in Natural Language Processing - Natural Language Processing in the Generative AI Era, RANLP 2025
Country/TerritoryBulgaria
CityVarna
Period8/09/2510/09/25

ASJC Scopus subject areas

  • Software
  • Computer Science Applications
  • Artificial Intelligence
  • Electrical and Electronic Engineering

Fingerprint

Dive into the research topics of 'Towards Safer Hebrew Communication: A Dataset for Offensive Language Detoxification'. Together they form a unique fingerprint.

Cite this