The Iranian linguistic family is pluricentric, encompassing Iranian Persian, Dari (Afghanistan), and Tajiki (Tajikistan). While Multilingual Large Language Models (MLLMs) claim broad coverage, their robustness across these regional variants and script differences (Perso-Arabic vs. Cyrillic) remains under-explored, particularly in the open-weight landscape. We evaluate five openweight models from the Qwen, Bloomz, and Gemma families across four downstream tasks: Sentiment Analysis, Machine Translation (MT), NLI, and QA. Utilizing a dataset of over 240,000 processed samples, we observe severe performance disparities. While the fine-tuned gemma-3-4b-persian achieves promising results on Iranian Persian (77.3% accuracy in Sentiment), almost all tested models appear to suffer catastrophic degradation on Tajiki script (dropping to 1.0 BLEU). These findings highlight a critical “script barrier” in current open-weight MLLM development for Central Asian languages. Code and data available here.

Gender bias continues to shape societal perceptions across both STEM (Science, Technology, Engineering, and Mathematics) and SHAPE (Social Sciences, Humanities, and the Arts for People and the Economy) domains. While existing studies have explored such biases in English language models, similar analyses in Bangla—spoken by over 240 million people—remain scarce. In this work, we investigate gender-profession associations in Bangla language models. We introduce Pokkhopat, a curated dataset of gendered terms and profession-related words across STEM and SHAPE disciplines. Using a suite of embedding-based bias detection methods—including WEAT, ECT, RND, RIPA, and cosine similarity visualizations—we evaluate 11 Bangla language models. Our findings show that several widely-used open-source Bangla NLP models (e.g., sagorsarker/bangla-bert-base) exhibit significant gender bias, underscoring the need for more inclusive and bias-aware development in low-resource languages like Bangla. We also find that many STEM and SHAPE-related words are absent from these models’ vocabularies, complicating bias detection and possibly amplifying existing biases. This emphasizes the importance of incorporating more diverse and comprehensive training data to mitigate such biases moving forward. Code available at https://github.com/HerWILL-Inc/ACL-2025/.

Misogynistic memes on social media perpetuate gender stereotypes, contribute to harassment, and suppress feminist activism. However, most existing misogyny detection models focus on high-resource languages, leaving a gap in low-resource settings. This work addresses that gap by focusing on misogynistic memes in Tamil and Malayalam, two Dravidian languages with limited resources. We combine computer vision and natural language processing for multi-modal detection, using CLIP embeddings for the vision component and BERT models trained on code-mixed hate speech datasets for the text component. Our results show that this integrated approach effectively captures the unique characteristics of misogynistic memes in these languages, achieving competitive performance with a Macro F1 Score of 0.7800 for the Tamil test set and 0.8748 for the Malayalam test set. These findings highlight the potential of multimodal models and the adaptation of pre-trained models to specific linguistic and cultural contexts, advancing misogyny detection in low-resource settings. Code available at https://github.com/HerWILL-Inc/NAACL-2025