Morocco

Large Language Models demonstrate strong safety behavior when user intent is expressed in English, yet their robustness under code switching and transliteration remains insufficiently understood. This paper investigates code switching as a safety failure mode, focusing on Roman Urdu, a widely used transliterated and mixed language form employed in informal and emotionally expressive communication. We introduce the Roman Urdu Adversarial Benchmark, a semantically controlled evaluation set comprising 400 distinct safety-relevant intents, each expressed in three aligned variants: English, mixed Roman Urdu with limited English usage, and purely transliterated Roman Urdu without English words, yielding 1,200 total evaluation inputs. The benchmark spans four critical safety categories: passive suicidal ideation, psychological distress, threat or intimidation, and coercion or emotional manipulation. We conduct a broad comparative evaluation of seven state-of-the-art language models from OpenAI, Anthropic, and xAI using recall-based safety metrics. Results reveal consistent degradation in safety detection under Roman Urdu inputs, with particularly pronounced drops for passive suicidal ideation. At the same time, instruction-tuned and reasoning-capable models demonstrate improved robustness, indicating that such failures are not inevitable and motivating transliteration-aware safety alignment methods.

EACL 2026

Code-Switching as a Safety Failure Mode in Large Language Models: An Empirical Study of Roman Urdu across English, Mixed, and Transliteration-Only Inputs

multilingual robustness

roman urdu

safety evaluation

llm safety

cross-lingual alignment

transliteration

code-switching

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)*<br><br>

<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>
<br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

workshop paper

Encoder-only transformer models remain widely used for discriminative NLP tasks, yet recent architectural advances have largely focused on English. In this work, we present \textbf{AraModernBERT}, an adaptation of the ModernBERT encoder architecture to Arabic, and study the impact of transtokenized embedding initialization and native long-context modeling up to 8,192 tokens. We show that transtokenization is essential for Arabic language modeling, yielding dramatic improvements in masked language modeling performance compared to non-transtokenized initialization. We further demonstrate that AraModernBERT supports stable and effective long-context modeling, achieving improved intrinsic language modeling performance at extended sequence lengths. Downstream evaluations on Arabic natural language understanding tasks, including inference, offensive language detection, question-question similarity, and named entity recognition, confirm strong transfer to discriminative and sequence labeling settings. Our results highlight practical considerations for adapting modern encoder architectures to Arabic and other languages written in Arabic-derived scripts.

AraModernBERT: Transtokenized Initialization and Long-Context Encoder Modeling for Arabic

Arabic speech recognition systems face distinct challenges due to the language's complex morphology and dialectal variations. Self-supervised models (SSL) like XLS-R have shown promising results, but their size with over than 300 million of parameters, makes fine-tuning computationally expensive. In this work, we present the first comparative study of parameter-efficient fine-tuning (PEFT), specifically LoRA and DoRA, applied to XLS-R for Arabic ASR. We evaluate on the newly released Common Voice Arabic V24.0 dataset, establishing new benchmarks. Our full fine-tuning achieves state-of-the-art results among XLS-R-based models with 23.03% Word Error Rate (WER). In our experiments, LoRA achieved a 36.10% word error rate (WER) while training just 2% of the model's parameters. DoRA reached 45.20% WER in initial experiments. We analyze the trade-offs between accuracy and efficiency, offering practical guidance for developing Arabic ASR systems when computational resources are limited. The models and code are publicly available.

Parameter-Efficient Adaptation of Self-Supervised Models for Arabic Speech Recognition

This work presents an evaluation of large language models (LLMs) for English to dialectal Arabic machine translation on the MADAR dataset. We evaluate both translation directions (English to Arabic and vice-versa) on 16 Arabic dialects. Our experiments cover a diverse set of models, including specialized Arabic models (Jais, Nile), multilingual models (Gemma, Command-R, Mistral, Aya), and commercial APIs (GPT-4.1). We employ multiple evaluation metrics: BLEU, CHRF, COMET (both reference-based and reference-less variants) and GEMBA (LLM-as-a-judge), as well as a small-scale manual evaluation, to assess translation quality. We discuss the challenges of automatic MT evaluation, especially in the context of Arabic dialects. We also evaluate the ability of LLMs to classify the dialect used in a text. The study offers insights into the capabilities and limitations of current LLMs for dialectal Arabic machine translation, particularly highlighting the difficulty of handling dialectal diversity.

Current state of LLMs for Arabic dialectal machine translation

We present AraLingBench, a fully human annotated benchmark for evaluating the Arabic linguistic competence of large language mod- els (LLMs). The benchmark spans five core categories: grammar, morphology, spelling, reading comprehension, and syntax, through 150 expert designed multiple choice questions that directly assess structural language understanding. Evaluating 35 Arabic and bilingual LLMs reveals that current models demonstrate strong surface level proficiency but struggle with deeper grammatical and syntactic reasoning. AraLingBench highlights a persistent gap between high scores on knowledge-based benchmarks and true linguistic mastery, showing that many models succeed through memorization or pattern recognition rather than au- thentic comprehension. By isolating and measuring fundamental linguistic skills, AraLingBench provides a diagnostic framework for developing Arabic LLMs

AraLingBench: A Human-Annotated Benchmark for Evaluating Arabic Linguistic Capabilities of Large Language Models

Arabic language faces technical and cultural challenges, including a lack of high-quality resources and the prevalence of regional dialects, which hinders the development of effective language processing systems. Therefore, the "Murabaa" platform was developed to transform Arabic linguistic knowledge into integrated digital resources. The platform aims to provide accurate digital content and promote the use of Arabic in various fields to bridge the gap between tradition and modernity by offering integrated linguistic resources for developing advanced research tools. The platform provides eight accurate dictionaries in the form of a website and a web application, contributing to the digitization of knowledge and its representation within the framework of standard lexical markup. In this study, we also conduct a quantitative comparison of the resources against similar ones to assess the quality of the linguistic knowledge they provide.

Murabaa: A comprehensiveResource Platform for Arabic Morphology

Medical AI systems increasingly rely on large language models (LLMs), yet their deployment in linguistically diverse regions remains unexplored. We address this gap by introducing U-MIRAGE, the first medical question-answering benchmark for Urdu and Roman Urdu. Urdu is the 11th most spoken language (with over 246 million speakers) worldwide. Our systematic evaluation of six state-of-the-art LLMs reveals three main findings. (1) Performance drops 6-10% points when moving from English to Urdu variants, even though medical knowledge should theoretically transfer across languages. (2) Chain-of-Thought (CoT) prompting improves small models by 8-20% points, while surprisingly the larger models' performance degraded by up to 3%. (3) Quantized small models fail catastrophically in low-resource languages, achieving near-random accuracy regardless of various prompting strategies. These findings challenge core assumptions about multilingual medical AI systems. Roman Urdu consistently outperforms standard Urdu script, suggesting orthographic alignment with pre-training data matters more than linguistic proximity. CoT prompting effectiveness depends critically on model architecture rather than task complexity alone. CoT prompting effectiveness depends critically on model architecture rather than task complexity alone. Our contributions are threefold: (1) U-MIRAGE, (2) systematic benchmarking of LLMs for Urdu and Roman Urdu medical reasoning, and (3) empirical analysis of CoT prompting in low-resource contexts. Our code and datasets are publicly available

U-MIRAGE: Benchmarking Chain-of-Thought Reasoning for Urdu Medical QA

Framing detection in Arabic social media is difficult due to interpretive ambiguity, cultural grounding, and limited reliable supervision. Existing LLM-based weak supervision methods typically rely on label aggregation, which is brittle when annotations are few and socially dependent. We propose a reliability-aware weak supervision framework that shifts the focus from label fusion to data curation. A small multi-agent LLM pipeline—two framers, a critic, and a discriminator—treats disagreement and reasoning quality as epistemic signals and produces instance-level reliability estimates. These estimates guide a QUBO-based subset selection procedure that enforces frame balance while reducing redundancy. Intrinsic diagnostics and an out-of-domain Arabic sentiment transfer test show that the selected subsets are more reliable and encode non-random, transferable structure, without degrading strong text-only baselines.

Optimizing What We Trust: Reliability-Guided QUBO Selection of Multi-Agent Weak Framing Signals for Arabic Sentiment Prediction

We introduce the Tarab Corpus, a large-scale cultural and linguistic resource that brings together Arabic song lyrics and poetry within a unified analytical framework. The corpus comprises 2.56 million verses and more than 13.5 million tokens, making it, to our knowledge, the largest open Arabic corpus of creative text spanning both classical and contemporary production. Tarab is broadly balanced between songs and poems and covers Classical Arabic, Modern Standard Arabic (MSA), and six major regional varieties: Egyptian, Gulf, Levantine, Iraqi, Sudanese, and Maghrebi Arabic. The artists and poets represented in the corpus are associated with 28 modern nation states and multiple historical eras, covering over fourteen centuries of Arabic creative expression from the Pre-Islamic period to the twenty-first century. Each verse is accompanied by structured metadata describing linguistic variety, geographic origin, and historical or cultural context, enabling comparative linguistic, stylistic, and diachronic analysis across genres and time. We describe the data collection, normalisation, and validation pipeline and present baseline analyses for variety identification and genre differentiation. The dataset will be released publicly on GitHub, with the access link omitted for anonymisation.

Tarab: A Multi-Dialect Corpus of Arabic Lyrics and Poetry

Context: Natural Language Processing (NLP) has become an essential field with widespread applications across domains such as Large Language Models (LLMs). One of the core applications of NLP is machine translation (MT). A major challenge in MT is handling out-of-vocabulary (OOV) words and spelling mistakes, which can lead to poor translation quality.
Objective: This study compares traditional text-based embeddings with visual embeddings for English-to-Arabic translation. It investigates the effectiveness of each approach, especially in handling noisy inputs or OOV terms.
Method: Using the IWSLT 2017 English-Arabic dataset, we trained a baseline transformer encoder-decoder model using standard text embeddings and compared it with models using several visual embeddings strategies, including vowel-removal preprocessing and trigram-based image rendering. The translated outputs were evaluated using BLEU scores.
Results: indicate that while traditional embeddings achieve higher BLEU scores on clean data, visual embeddings demonstrate superior robustness to input noise and OOV terms, particularly under high-noise conditions.

Seeing Words Differently: Visual Embeddings for Robust English-Arabic Machine Translation

Arabic dialect↔English machine translation remains difficult due to extreme dialect variation, inconsistent orthography, and limited parallel data. Moreover, dialect translation is often needed in remote regions or by economically-disadvantaged communities, which often operate in compute-constrained or offline settings. Motivated by these concerns, in this paper we explore optimizing Arabic dialect↔English translators that run over small LLMs, which could be implemented on small offline devices.

We show that reasoning-oriented reinforcement learning can substantially improve small multilingual LLMs for Arabic dialect translation. Using the MADAR corpus, small Qwen-2.5 models trained with a think-then-translate template and optimized with Group-Relative Policy Optimization using a SacreBLEU reward outperform a much larger 7B baseline trained with supervised fine-tuning. The dialect-to-English BLEU score more than doubles from 17.4 to 34.9, while the English-to-dialect COMET score improves from 0.57 to 0.73.

Downloads

Next from EACL 2026

AraModernBERT: Transtokenized Initialization and Long-Context Encoder Modeling for Arabic

Similar lecture

U-MIRAGE: Benchmarking Chain-of-Thought Reasoning for Urdu Medical QA

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from EACL 2026

AraModernBERT: Transtokenized Initialization and Long-Context Encoder Modeling for Arabic

Similar lecture

U-MIRAGE: Benchmarking Chain-of-Thought Reasoning for Urdu Medical QA

Downloads