arXiv AI

'Ghaib in Translation' aka Unseen Harm: Measuring Cross-Script Safety Inconsistency with 'Missed-in-Urdu' Scores in LLM Hate Speech Detection

The paper "Ghaib in Translation" investigates how large language models (LLMs) handle Urdu, a widely spoken language that is largely absent from safety evaluations. Five prominent LLMs—GPT‑4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen‑2.5, and Llama‑3.1—were tested on six datasets covering Nastaliq Urdu, Roman Urdu, English, and code‑switched Urdu‑English. The study found significant label instability between original‑script and English‑translation classifications, with missed‑in‑Urdu rates ranging from 2.4% to 9.9% (median 4.3%). A review of 205 papers across nine ALW/WOAH editions revealed no dedicated Urdu research, underscoring the language’s neglect in current safety research.

arXiv AI
Aug 20

Efficient Adaptation of LLMs for Hate Speech Detection in Low-Resource Languages: A Comparative Study on Roman Urdu

The paper evaluates large language models for hate‑speech detection in Roman Urdu, a low‑resource language with informal spelling variations. Using the Parameter‑Efficient Fine‑Tuning technique Low‑Rank Adaptation (LoRA), the authors fine‑tune models such as Mistral, LLaMA, Falcon, and multilingual BERT on the 72,000‑comment PURUTT dataset. While zero‑shot performance yields an F1 of 0.56, fine‑tuning a small fraction of parameters boosts F1 scores above 0.93, demonstrating that PEFT offers both high accuracy and computational efficiency for low‑resource language tasks.

By Toneema Zubair, Muhammad Junaid Asif, Faisal Kamiran, Hafiz Hassan Saeed, Rana Fayyaz Ahmad
arXiv AI
5d ago

Register Shifts Break LLM Safety: A Bengali Benchmark with Culturally Grounded Harms

The paper introduces BanglaSafe, a benchmark of 879 Bengali prompts that covers 17 culturally grounded harm categories and five prompting conditions. Evaluation of 18 frontier LLMs shows that 53.6% of responses are unsafe or partially unsafe, with 14.7% containing strictly harmful content. The study finds that the writing style within Bengali has a stronger impact on safety than the language switch itself, and that current safety classifiers struggle to reliably evaluate Bengali content.

By Naymul Islam, Nusrat Jahan Lia, Shubhashis Roy Dipta, Sabik Bin Sultan, Abdullah Khan Zehady
arXiv AI
Aug 20

Safety Alignment Illusion: The Cross-Lingual Safety Gap in LLMs

The paper titled "Safety Alignment Illusion: The Cross-Lingual Safety Gap in LLMs" highlights that current safety alignment training for large language models is predominantly English-centric, leading to failures in non‑English languages. It introduces INCLUDE, a multilingual benchmark with 2,604 prompts in six languages (English, Hindi, Bengali, Marathi, Tamil, and Hinglish) to measure Indian‑centric socio‑cultural biases. Evaluation of ten open‑ and closed‑source LLMs shows that Bengali models exhibit the highest bias scores among open‑source models, while English shows the lowest bias in open‑source but the highest in closed‑source models.

By Namya Bhatnagar
arXiv Machine Learning
4d ago

Contextual Embedding Evidence for Main--Light Verb Distinctions in Urdu

The study examines Urdu light verbs, which add schematic event meaning while staying lexically linked to their main verbs. Using contextual embeddings from UrduBERT, DunbaaBERT, and multilingual BERT on 1,126 sentences, the authors find significant representational separation between main and light uses across all verb–model pairs, yet main and light uses of the same lemma remain closer than mismatched pairs. In a seven‑way prediction task limited to light uses, UrduBERT achieves 0.866 accuracy and 0.852 macro‑F1, and maintains 0.782 accuracy when tested on unseen preceding forms, demonstrating generalization beyond local verb combinations.

By Farah Adeeba, Miriam Butt
arXiv AI
Jun 9

Culturally-Adapted Red-Teaming Across East and Southeast Asian Contexts: A Methodological and Comparative Analysis

arXiv:2606. 09178v1 Announce Type: cross Abstract: Multilingual safety evaluation of large language models (LLMs) has predominantly relied on direct translation (DT) of English benchmarks into target languages - an approach that converts surface-level linguistic form while failing to reflect the cultural context embedded in threat scenarios, social norms, and legal frameworks.

By Hyeji Choi, Yongtaek Lim, Minwoo Kim
Hugging Face Trending Papers
Jun 8

Culturally-Adapted Red-Teaming Across East and Southeast Asian Contexts: A Methodological and Comparative Analysis

Multilingual safety evaluation of large language models (LLMs) has predominantly relied on direct translation (DT) of English benchmarks into target languages - an approach that converts surface-level linguistic form while failing to reflect the cultural context embedded in threat scenarios, social norms, and legal frameworks. We construct paired DT and culturally-adapted (CA) datasets via 1:1 seed matching for four languages - Korean (KO), Japanese (JA), Thai (TH), and Khmer (KM) - and compare Attack Success Rate (ASR) and Cultural Realism scores across four open-source LLM.