We study the problem of preserving offensive load in LLM-generated translations, focusing on how offensive words retain their load across languages. We draw on HurtNet, a novel language resource consisting of five multilingual (Arabic, French, Bulgarian, Greek, and Italian) dictionaries of hurtful terms, with multiple definitions and examples of usage. We draw on HurtNet to identify offensive senses shared by the 5 languages and the terms that denote them, and feed translations with this knowledge. We evaluate the approach through quantitative and qualitative analysis. A series of knowledge-injection zero-shot translation experiments were carried out with four different prompts that gradually increase injected linguistic knowledge, and three LLMs —an open use, a translation-dedicated, and a commercial one. Our results show that translations better preserving offensive load are produced when the prompt explicitly instructs the model about the translated offensive term and that the method is beneficial for those cases where the literal translation of the term does not preserve its offensive load.

This paper presents a methodology that uses LLMs to align multilingual offensive lexicons at the sense level. Lexicons of different structures and origins in Arabic, Bulgarian, Modern Greek, French, and Italian have been aligned directly without pivoting through English. The Modern Greek lexicon is LLM-generated, and the other four lexicons are WordNet-compatible. For inter-language alignment of senses, an LLM-as-a-judge rubric was used over lemma–definition–example triples. The LLM makes 2.87M pairwise comparisons and yields 31 strict global-sense categories. The paper discusses the challenges involved in sense alignment tasks. The resource is available to support downstream applications such as Machine Translation and cross-lingual hate-speech detection.