Elamur Research⌕
☰
home / papers / 2026 / From Form to Meaning: Interlingua Sense-Alignment of Offensive Language with LLMs

From Form to Meaning: Interlingua Sense-Alignment of Offensive Language with LLMs

Summary

Offensive words rarely translate cleanly. The same form can carry different insulting meanings across cultures, and routing everything through English tends to erase these differences. For example, Greek "γαϊδούρι" ('donkey') implies rudeness rather than stupidity. The authors propose an LLM-based way to link offensive lexicons sense by sense, working directly between languages and never through an English pivot. They use the Hurt-Net collection, covering Arabic, Bulgarian, French, Italian and Modern Greek. Greek had no suitable sense-separated offensive dictionary, so they built one with an LLM, recording definitions, usage examples and target classes for each sense. An LLM-as-a-judge rubric then compares lemma–definition–example triples in their original languages. This yields global sense categories shared across all five languages, along with senses that resist alignment because they are tied to one culture or register. The resulting resource is meant to support offense-aware Machine Translation and cross-lingual hate-speech detection. The authors identify lexical gaps, mismatched sense granularity and partial overlaps between meanings as the main obstacles.

Key points

  • Scale: the pipeline ran 2.87M pairwise sense comparisons and produced 31 strict global-sense categories. The four non-Greek lexicons are WordNet-compatible.
  • The Greek lexicon started from HurtLex-EL. The prompt was developed with GPT-o3 over 15 iterations on 37 pilot lemmas. It requires a genus–differentia definition for each sense, three GDEX-style examples per sense, and one of six CIDOC CRM–based target classes (Person, Object, State, Repetitive Action, Event, Behavioral Trait).
  • Some senses could not be merged with any counterpart, such as French trou ('prison'), Bulgarian шваба (a slur for Germans) and Greek αμερικανάκι.
  • Problems found in the LLM-generated Greek content include archaic forms, interference from English, evaluative tone leaking into definitions, and semantic drift in the examples.
  • Bilingual human evaluation covered only some language pairs. The code and prompts are available on GitHub (aroussopoulou/Interlingua-Sense-Alignment-of-Offensive-Language-with-LLMs).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Machine translation · Natural language processing · Society & ethics

Topics: Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.ltedi-1.6

@inproceedings{roussopoulou2026form,
  title = {{From Form to Meaning: Interlingua Sense-Alignment of Offensive Language with LLMs}},
  author = {Roussopoulou, Maria Alexandra and Markantonatou, Stella},
  booktitle = {Proceedings of the Sixth Workshop on Language Technology for Equality, Diversity, Inclusion},
  year = {2026},
  doi = {10.18653/v1/2026.ltedi-1.6},
  url = {https://doi.org/10.18653/v1/2026.ltedi-1.6}
}

References

66 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Elisabeth Camp. 2013. Slurring perspectives. Analytic Philosophy, 54(3):330–349. find
  2. Sai Saketh Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020. Deep learning models for multilingual hate speech detection. Preprint, arXiv:2004.06465. ↗
  3. Centre for the Greek Language. Portal for the greek language. Website. Accessed: 2026-02-25. find
  4. Fai Leui Chan, Duke Nguyen, and Aditya Joshi. 2024. “is hate lost in translation?”: Evaluation of multilingual LGBTQIA+ hate speech detection. In Proceedings of the 22nd Annual Workshop of the Australasian Language Technology Association, pages 146–152, Canberra, Australia. Association for Computational Linguistics. find
  5. B. T. S. Atkins and M. Rundell. 2008. The Oxford Guide to Practical Lexicography. Oxford University Press, Oxford. ↗
  6. George Babiniotis. 2024. Dictionary of Modern Greek, 6th edition. Lexicology Centre. Accessed: 2026-02-25. find
  7. Yi-Ling Chung, Elizaveta Kuzmenko, Serra Sinem Tekiroglu, and Marco Guerini. 2019. CONAN -COunter NArratives through nichesourcing: a multilingual dataset of responses to fight online hate speech. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2819–2829, Florence, Italy. Association for Computational Linguistics. find our card
  8. Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86–90, Montreal, Quebec, Canada. Association for Computational Linguistics. ↗ our card
  9. CIDOC CRM Special Interest Group. 2021. CIDOC Conceptual Reference Model (CRM), Version 7.2. ISO 21127:2023 correspondence document. find
  10. Miloš Jakubíček and Michael Rundell. 2023. The end of lexicography? can ChatGPT outperform current tools for post-editing lexicography? In Proceedings of eLex 2023. find
  11. Philipp Cimiano, John P. McCrae, Paul Buitelaar, and 1 others. 2016. Lexicon model for ontologies: Ontolex-lemon. W3c community group report, W3C Ontology Lexicon Community Group. Published 10 May 2016. find
  12. Aiqi Jiang and Arkaitz Zubiaga. 2021. Cross-lingual capsule network for hate speech detection in social media. In Proceedings of the 32nd ACM Conference on Hypertext and Social Media, HT ’21, page 217–223, New York, NY, USA. Association for Computing Machinery. find
  13. Jonathan Culpeper. 2011. Impoliteness metadiscourse, page 71–112. Studies in Interactional Sociolinguistics. Cambridge University Press. find
  14. Temuulen Khishigsuren, Gábor Bella, Khuyagbaatar Batsuren, Abed Alhakim Freihat, Nandu Chandran Nair, Amarsanaa Ganbold, Hadi Khalilia, Yamini Chandrashekar, and Fausto Giunchiglia. 2022. Using linguistic typology to enrich multilingual lexicons: the case of lexical gaps in kinship. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2798–2807, Marseille, France. European Language Resources Association. find
  15. Gilles-Maurice de Schryver. 2024. The road towards fine-tuned LLMs for lexicography. In Book of Abstracts of the Workshop “Large Language Models and Lexicography” (EURALEX 2024 workshop), pages 6–11, Ljubljana. ELEXIS Association. find
  16. Alphaeus Dmonte, Shrey Satapara, Rehab Alsudais, Tharindu Ranasinghe, and Marcos Zampieri. 2024. On the effects of machine translation on offensive language detection. Social Network Analysis and Mining, 14(1):242. ↗
  17. Adam Kilgarriff, Miloš Husák, Katy McAdam, Michael Rundell, and Pavel Rychlý. 2008. GDEX: Automatically finding good dictionary examples in a corpus. In Proceedings of the 13th EURALEX International Congress, pages 425–432, Barcelona, Spain. find our card
  18. Stephanie Evert, Christine Ganslmayer, and Christian Rink. 2024. Multi-level analysis as a systematic approach to evaluating the quality of ai-generated dictionary entries. In Lexicography and Semantics. Proceedings of the XXI EURALEX International Congress, pages 317–335, Cavtat. Institut za hrvatski jezik. find
  19. S.V. Kogilavani, S. Malliga, K.R. Jaiabinaya, M. Malini, and M. Manisha Kokila. 2023. Characterization and mechanical properties of offensive language taxonomy and detection techniques. Materials Today: Proceedings, 81:630–633. International Virtual Conference on Sustainable Materials (IVCSM-2k20). find
  20. Mariia Fedorova, Andrey Kutuzov, and Yves Scherrer. 2024. Definition generation for lexical semantic change detection. In Findings of the Association for Computational Linguistics: ACL 2024, pages 5712– 5724, Bangkok, Thailand. Association for Computational Linguistics. find
  21. Katerina Korre, Arianna Muti, and Alberto Barrón-Cedeño. 2024. The challenges of creating a parallel multilingual hate speech corpus: An exploration. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15842–15853, Torino, Italia. ELRA and ICCL. find
  22. Ine Gevers, Ilia Markov, and Walter Daelemans. 2022. Linguistic analysis of toxic language on social media. Computational Linguistics in the Netherlands Journal, 12:33–48. find
  23. Katerina Korre, Arianna Muti, Federico Ruggeri, and Alberto Barrón-Cedeño. 2025. Untangling hate speech definitions: A semantic componential analysis across cultures and domains. Preprint, arXiv:2411.07417. ↗
  24. Erfan Ghadery and Marie-Francine Moens. 2020. LIIR at SemEval-2020 task 12: A cross-lingual augmentation approach for multilingual offensive language identification. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 2073–2079, Barcelona (online). International Committee for Computational Linguistics. find
  25. Senyu Li, Bradley Hauer, Ning Shi, and Grzegorz Kondrak. 2024. Translation-based lexicalization generation and lexical gap detection: Application to kinship terms. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6891–6900, Bangkok, Thailand. Association for Computational Linguistics. find
  26. Yi Han, Ryohei Sasano, and Koichi Takeda. 2024. Definition generation for automatically induced semantic frame. In Findings of the Association for Computational Linguistics: ACL 2024, pages 11112–11118, Bangkok, Thailand. Association for Computational Linguistics. find
  27. LOD Cloud. Open multilingual wordnet (greek) dataset (omwn-ell). Dataset metadata page. Accessed: 2026-03-02. find
  28. Muhammad Okky Ibrohim and Indra Budi. 2019. Translated vs non-translated method for multilingual hate speech identification in twitter. International Journal of Advanced Science, Engineering and Information Technology, 9(4):1116–1123. ↗
  29. Sebastian Loftus, Adrian Mülthaler, Sanne Hoeken, Sina Zarrieß, and Ozge Alacam. 2025. Using LLMs and preference optimization for agreement-aware HateWiC classification. In Proceedings of the The 9th Workshop on Online Abuse and Harms (WOAH), pages 538–547, Vienna, Austria. Association for Computational Linguistics. find
  30. Endang Wahyu Pamungkas, Valerio Basile, and Viviana Patti. 2021. A joint learning approach with knowledge injection for zero-shot cross-lingual hate speech detection. Inf. Process. Manag., 58(4):102544. {Simo K.} Määttä. 2023. Linguistic and discursive properties of hate speech and speech facilitating the expression of hatred: Evidence from finnish and french online discussion boards. Internet Pragmatics, 6(2):156–172. ↗
  31. Endang Wahyu Pamungkas, Valerio Basile, and Viviana Patti. 2023. Towards multidomain and multilingual abusive language detection: a survey. Personal and Ubiquitous Computing, 27:17–43. find
  32. Stella Markantonatou, Vivian Stamou, Christina Christodoulou, Georgia Apostolopoulou, Antonis Balas, and George Ioannakis. 2024. Aikia corpus. OSF repository. find
  33. Francesco Periti, David Alfter, and Nina Tahmasebi. 2024. Automatically generated definitions and their utility for modeling word meaning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14008–14026, Miami, Florida, USA. Association for Computational Linguistics. find
  34. Domenico Meconi, Simone Stirpe, Federico Martelli, Leonardo Lavalle, and Roberto Navigli. 2025. Do large language models understand word senses? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 33897–33916, Suzhou, China. Association for Computational Linguistics. ↗ our card
  35. Bach Pham, JuiHsuan Wong, Samuel Kim, Yunting Yin, and Steven Skiena. 2025. Word definitions from large language models. In 2025 19th International Conference on Semantic Computing (ICSC), page 158–162. IEEE. find
  36. Mistral AI. 2025. Mistral small 3.2 (v25.06) | mistral docs. https://docs.mistral.ai/models/mis tral-small-3-2-25-06. Published: 2025-06-20. Accessed: 2026-02-26. ↗
  37. Chayanon Phoodai and Richárd Rikk. 2023. Exploring the capabilities of ChatGPT for lexicographical purposes: A comparison with Oxford Advanced Learner’s Dictionary within the microstructural framework. In Proceedings of eLex 2023. find
  38. Mistral AI. 2026a. Batch inference | mistral docs. https://docs.mistral.ai/capabilities/batch/. Accessed: 2026-02-26. ↗
  39. Wessel Poelman and Miryam de Lhoneux. 2025. The roles of English in evaluating multilingual language models. In Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies (NoDaLiDa/Baltic-HLT 2025), pages 492–498, Tallinn, Estonia. University of Tartu Library. ↗
  40. Mistral AI. 2026b. Usage (chat completions) | mistral docs. https://docs.mistral.ai/capabilities/completion/usage. Accessed: 2026-02-26. ↗
  41. Khaled Mnassri, Reza Farahbakhsh, Rasoul Chalehchaleh, Pramudith Rajapaksha, Amir Reza Jafari, Guoliang Li, and Noel Crespi. 2024. A survey on multi-lingual offensive language detection. PeerJ Computer Science, 10:e1934. ↗
  42. Tharindu Ranasinghe and Marcos Zampieri. 2020. Multilingual offensive language identification with cross-lingual embeddings. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5838–5844, Online. Association for Computational Linguistics. find
  43. Roberto Navigli and Simone Paolo Ponzetto. 2010. BabelNet: Building a very large multilingual semantic network. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 216–225, Uppsala, Sweden. Association for Computational Linguistics. find
  44. Tharindu Ranasinghe and Marcos Zampieri. 2021a. An evaluation of multilingual offensive language identification methods for the languages of india. Information, 12(8). : © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/ 4.0/). ↗
  45. OpenAI. 2025. Introducing o3 and o4-mini. OpenAI Blog. find
  46. OpenAI. n.d. Pricing | openai api. https://develo pers.openai.com/api/docs/pricing. Accessed: 2026-02-25. ↗
  47. Tharindu Ranasinghe and Marcos Zampieri. 2021b. Multilingual offensive language identification for low-resource languages. Preprint, arXiv:2105.05996. ↗
  48. Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4675– 4684, Hong Kong, China. Association for Computational Linguistics. find our card
  49. Paul Röttger, Haitham Seelawi, Debora Nozza, Zeerak Talat, and Bertie Vidgen. 2022. Multilingual Hate-Check: Functional tests for multilingual hate speech detection models. In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH), pages 154–169, Seattle, Washington (Hybrid). Association for Computational Linguistics. find
  50. Happy Khairunnisa Sariyanto, Diclehan Ulucan, Oguzhan Ulucan, and Marc Ebner. 2025. Towards explainable hate speech detection. In Findings of the Association for Computational Linguistics: ACL 2025, pages 12883–12893, Vienna, Austria. Association for Computational Linguistics. find
  51. Piek Vossen, Francis Bond, and John P. McCrae. 2016. Toward a truly multilingual GlobalWordnet grid. In Proceedings of the 8th Global WordNet Conference (GWC), pages 424–431, Bucharest, Romania. Global Wordnet Association. find
  52. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. Preprint, arXiv:2201.11903. ↗ our card
  53. Gilles Sérasset. 2015. Dbnary: Wiktionary as a lemon based rdf multilingual lexical resource. Semantic Web. Dataset Description paper (Semantic Web Journal). find
  54. Gilles Sérasset. 2025. Towards sense to sense linking across DBnary languages. In Proceedings of the 5th Conference on Language, Data and Knowledge, pages 318–327, Naples, Italy. Unior Press. find
  55. Wiktionary contributors. Wiktionary: The free dictionary. Website. Accessed: 2026-02-25. find
  56. Yi Xu, Laura Ruis, Tim Rocktäschel, and Robert Kirk. 2025. Investigating non-transitivity in llm-as-a-judge. Preprint, arXiv:2502.14074. ↗
  57. Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, Weicheng Ma, and Soroush Vosoughi. 2025. Judging the judges: A systematic study of position bias in LLM-as-a-judge. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 292–314, Mumbai, India. The Asian Federation of Natural Language Processing and The Association for Computational Linguistics. ↗
  58. George J. Xydopoulos, Anna Iordanidou, and Anastasia Efthymiou. 2009. Recent advances in the documentation of greek slang: The case of www.slang.gr. In Proceedings of the 9th International Conference on Greek Linguistics (ICGL9), Chicago, USA. Paper first presented at ICGL9 (October 2009); PDF distributed via OSU ICGL proceedings. find
  59. Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, Nitesh Chawla, and Xiangliang Zhang. 2025. Justice or prejudice? quantifying biases in llm-as-a-judge. In International Conference on Learning Representations, volume 2025, pages 102351–102390. ↗ our card
  60. Vivian Stamou, Iakovi Alexiou, Antigone Klimi, Eleftheria Molou, Alexandra Saivanidou, and Stella Markantonatou. 2022. Cleansing & expanding the HURTLEX(el) with a multidimensional categorization of offensive words. In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH), pages 102–108, Seattle, Washington (Hybrid). Association for Computational Linguistics. find
  61. Marcos Zampieri, Preslav Nakov, Sara Rosenthal, Pepa Atanasova, Georgi Karadzhov, Hamdy Mubarak, Leon Derczynski, Zeses Pitenis, and Çağrı Çöltekin. 2020. SemEval-2020 task 12: Multilingual offensive language identification in social media (OffensEval 2020). In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 1425–1447, Barcelona (online). International Committee for Computational Linguistics. ↗
  62. Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. 2024. Large language models are inconsistent and biased evaluators. Preprint, arXiv:2405.01724. ↗
  63. Simone Sulpizio, Fritz Günther, Linda Badan, Benjamin Basclain, Marc Brysbaert, Yuen Lai Chan, and 1 others. 2024. Taboo language across the globe: A multi-lab study. Behavior Research Methods, 56:3794– 3813. find
  64. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and chatbot arena. ArXiv:2306.05685v4 (NeurIPS 2023 Datasets and Benchmarks Track). ↗
  65. Hristo Tanev. 2024. JRC at ClimateActivism 2024: Lexicon-based detection of hate speech. In Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE 2024), pages 85–88, St. Julians, Malta. Association for Computational Linguistics. find
  66. A Muhammad Usman, Muhammad Ahmad, M. Shahiki Tash, Irina Gelbukh, Rolando Quintero Tellez, and Grigori Sidorov. 2025. Multilingual hate speech detection in social media using translation-based approaches with large language models. Preprint, arXiv:2506.08147. Judging Prompt You are an expert in offensive language and multilingual lexicography. Your task: given TWO OF-FENSIVE TERMS (Term A and Term B), decide how similar they are in sense of their offensive meaning and use. You must choose exactly ONE of: Bertie Vidgen and Leon Derczynski. 2020. Directions in abusive language training data, a systematic review: Garbage in, garbage out. PLOS ONE, 15(12):e0243300. • merge → same offensive sense (one Sense) • related → clearly related / near-synonyms, but not exactly the same sense • unrelated → no useful offensive relation Guidelines (short): • Use merge only if they share: 1. same offensive purpose (e.g., expletive for a bad event, insult for the same type of person), typical target (situa-2. same tion/person/group), 3. same core offensive idea and similar strength. • Use related if they are in the same offensive area but differ in purpose, target, or strength. • Use unrelated if their offensive domain or purpose is different, or if any link is too vague. Be conservative: • If between merge/related → choose related. • If between related/unrelated → choose unrelated. OUTPUT FORMAT (VERY IMPORTANT): Answer with EXACTLY ONE LINE, with 5 fields separated by |||, in this order: decision|||confidence|||globalsense_id||| globalsense_label_en|||rationale_en Where: • decision: one of merge, related, unrelated (lowercase) • confidence: a number between 0 and 1 with at most 2 decimals (e.g., 0.87) • sense_label_en: – if decision is merge or related: short English label (max 12 words) – if decision is unrelated: write null • rationale_en: 1–2 short sentences in English explaining your decision (mention purpose, target, and core offensive idea; do NOT print any slurs) Your entire reply MUST be this single line. Do NOT add explanations before or after. Do NOT use markdown or code fences. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/roussopoulou-2026-form-meaning-interlingua-sense-alignment/