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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, short

Summary

A methodology utilizing Large Language Models (LLMs) is presented for aligning multilingual offensive lexicons at the sense level. This approach directly aligns lexicons from Arabic, Bulgarian, Modern Greek, French, and Italian, bypassing the need to pivot through English. The Modern Greek lexicon was LLM-generated, while the other four are WordNet-compatible. An "LLM-as-a-judge" rubric, applied to lemma–definition–example triples, facilitated inter-language sense alignment. The LLM performed 2.87M pairwise comparisons, yielding 31 strict global-sense categories. The work also addresses challenges in sense alignment tasks, providing a valuable resource for downstream applications like Machine Translation and cross-lingual hate-speech detection.

Key takeaway

For NLP Engineers developing cross-lingual applications, this methodology offers a direct path to align offensive language lexicons. You can bypass English-centric pivoting, improving accuracy and cultural nuance in systems like hate-speech detection and machine translation. Consider adopting an LLM-as-a-judge approach for sense alignment to build more robust and culturally sensitive multilingual resources.

Key insights

LLMs can directly align multilingual offensive lexicons at the sense level without English pivoting.

Principles

• Direct inter-language alignment avoids English pivot bias.
• LLMs can serve as judges for sense alignment tasks.
• Sense-level alignment is crucial for offensive language.

Method

An LLM-as-a-judge rubric evaluates lemma–definition–example triples for 2.87M pairwise comparisons, yielding 31 global-sense categories across five languages.

In practice

• Supports cross-lingual hate-speech detection.
• Enhances Machine Translation quality.
• Generates lexicons for low-resource languages.

Topics

• Offensive Language Detection
• Multilingual Lexicons
• Large Language Models
• Sense Alignment
• Cross-lingual NLP
• Hate Speech Detection

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• G-IdiomAlign: A Gloss-Pivoted Benchmark for Cross-Lingual Idiom Alignment — Cross-lingual idiom alignment challenges LLMs due to non-compositionality, with glosses offering semantic grounding but literal bias remaining dominant.
• Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment — A plug-and-play MLC loss improves multilingual LLM safety alignment using only prompt variants, reducing resource needs.
• Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection — A mechanism-oriented taxonomy of indirect linguistic encoding significantly enhances LLM-based coded language detection.
• The Wrong Kind of Right: Quantifying and Localizing Misfired Alignment in LLMs — Current LLM alignment can overgeneralize safety cues, overriding explicit evidence in stereotype-sensitive contexts.
• Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models — Automated metrics can identify and quantify lexical overuse and preference learning shifts in LLMs without manual curation.
• Linguistic Issues in Computational Linguistics for Non-English Languages — English-centric LLM design creates systematic linguistic weaknesses for non-English languages across multiple levels.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.