intella publishes two Arabic AI papers in a month

Arabic NLP isn’t just technology, it’s an opportunity to empower 400M voices and preserve a civilization’s knowledge.🌍Thrilled to share that 2 of my papers were accepted to the Arabic NLP Conference co-located with EMNLP this November in China: 1. Multimodal Arabic Captioning with Interpretable Visual Concept Integration: A two-stage Arabic captioning pipeline: retrieve interpretable Arabic visual labels (mCLIP, AraCLIP, Jina V4) from a ~21K concept vocabulary, then generate captions with Qwen-VL or Gemini Pro Vision. Our runs achieved Top 1 Cosine Similarity score, Top 2 LLM-as-a-Judge score🏅 🎉 https://lnkd.in/dK2W_yT4 2. Lexicon-Enriched Graph Modeling for Arabic Document Readability Prediction: A heterogeneous sentence–lemma graph infused with the SAMER readability lexicon complements a transformer sentence encoder. We combine both via late fusion and prioritize the hardest spans, yielding interpretable predictions grounded in lexicon cues and sentence relations—useful for education, curriculum design, and accessible Arabic reading tools. https://lnkd.in/dqKESBr4 Dr. Mervat Abu-Elkheir, Mayar Osama, Mohamad Rageh, Dr. Amany Fashwan My deepest thanks to the research partners—your dedication, creativity, and resilience made this possible. I’m also grateful to everyone who supported us; without your effort, none of this would have happened. 🙏 I won’t be able to travel to China and attend EMNLP in person this year, but I’ll be participating remotely—let’s connect online! #ArabicNLP #EMNLP2025 #ArabicAI #ImageCaptioning #GNN #Readability

Call for Papers is now open for SilkRoadNLP 2026! We’re thrilled to announce the 1st ACL workshop on NLP and Large Language Models for the Iranian linguistic family, co-located with EACL 2026 in Rabat, Morocco. With this workshop, we aim to bring together researchers, linguists, and AI practitioners to explore how multilingual and low-resource NLP can better represent Persian, Kurdish, Pashto, Balochi, Luri, Ossetic, and other Iranian languages. This initiative is co-organized by our team at the Zoorna Institute for Language, AI, and Society, as part of our mission to advance culturally aware, linguistically informed AI for low-resource languages. 🗓️ Submission deadline: 8 January 2026 🔗 silkroadnlp.org

𝐂𝐚𝐧 𝐰𝐞 𝐭𝐫𝐮𝐬𝐭 𝐠𝐥𝐨𝐛𝐚𝐥 𝐀𝐈 𝐦𝐨𝐝𝐞𝐥𝐬 𝐭𝐨 𝐫𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭 𝐀𝐫𝐚𝐛𝐢𝐜 𝐜𝐮𝐥𝐭𝐮𝐫𝐞 𝐟𝐚𝐢𝐫𝐥𝐲 𝐢𝐟 𝐀𝐫𝐚𝐛𝐢𝐜 𝐝𝐚𝐭𝐚 𝐢𝐬 𝐦𝐢𝐬𝐬𝐢𝐧𝐠? AI models are only as good as the data they learn from. Most large language models are trained primarily on English and a few other widely represented languages. Arabic ,with its rich history, dialects, and literary heritage, is often underrepresented. 𝐖𝐡𝐚𝐭 𝐝𝐨𝐞𝐬 𝐭𝐡𝐢𝐬 𝐦𝐞𝐚𝐧? 1) Cultural nuances, idioms, and traditional expressions might be misunderstood or ignored. 2) Poetry, literature, and historical texts could be misinterpreted. 3) Decisions driven by AI might reflect biases or gaps in understanding. In short, an AI trained without sufficient Arabic context is 𝐧𝐨𝐭 𝐧𝐞𝐮𝐭𝐫𝐚𝐥 ;𝐈𝐭 𝐫𝐢𝐬𝐤𝐬 𝐞𝐫𝐚𝐬𝐢𝐧𝐠 𝐨𝐫 𝐦𝐢𝐬𝐫𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐢𝐧𝐠 𝐚𝐬𝐩𝐞𝐜𝐭𝐬 𝐨𝐟 𝐨𝐮𝐫 𝐜𝐮𝐥𝐭𝐮𝐫𝐞. For Arabic NLP, this isn’t just a technical challenge. 𝐈𝐭’𝐬 𝐚 𝐫𝐞𝐬𝐩𝐨𝐧𝐬𝐢𝐛𝐢𝐥𝐢𝐭𝐲. Every dataset we curate, every model we train, is a 𝐬𝐭𝐞𝐩 𝐭𝐨𝐰𝐚𝐫𝐝 𝐞𝐧𝐬𝐮𝐫𝐢𝐧𝐠 𝐭𝐡𝐚𝐭 𝐭𝐞𝐜𝐡𝐧𝐨𝐥𝐨𝐠𝐲 𝐫𝐞𝐬𝐩𝐞𝐜𝐭𝐬 𝐚𝐧𝐝 𝐚𝐦𝐩𝐥𝐢𝐟𝐢𝐞𝐬 𝐀𝐫𝐚𝐛𝐢𝐜 𝐜𝐮𝐥𝐭𝐮𝐫𝐞, 𝐫𝐚𝐭𝐡𝐞𝐫 𝐭𝐡𝐚𝐧 𝐭𝐰𝐢𝐬𝐭𝐢𝐧𝐠 𝐚𝐧𝐝 𝐦𝐢𝐬𝐥𝐞𝐚𝐝𝐢𝐧𝐠 𝐢𝐭. As AI becomes part of our daily lives, we must ask: 𝐀𝐫𝐞 𝐰𝐞 𝐭𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝐢𝐭 𝐭𝐨 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝 𝐮𝐬, 𝐨𝐫 𝐚𝐫𝐞 𝐰𝐞 𝐥𝐞𝐭𝐭𝐢𝐧𝐠 𝐢𝐭 𝐝𝐞𝐟𝐢𝐧𝐞 𝐮𝐬? #ArabicNLP #AI #ArtificialIntelligence #LanguagePreservation #CulturalIdentity #NaturalLanguageProcessing #MachineLearning #ArabicLanguage #DigitalCulture #AIResearch #LanguageTech #TechForGood

ADAPT researcher Keren Artiaga from Munster Technological University (MTU) recently had a paper accepted to #EMNLP Findings. The study, titled “Rethinking Sign Language Translation: The Impact of Signer Dependence on Model Evaluation”, will be presented at the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), taking place from 4th to 9th November in Suzhou, China. The research highlights that performance reported in existing works does not reliably indicate how well sign language translation models truly generalise. Learn more: https://lnkd.in/evrPTNxk

Low-resource languages - Albanian Language 💗 Despite the remarkable progress in LLMs and pretrained models, our recent experiment on Albanian (Text-to-Speech) TTS and (Speech-to-Text) STT reveals a critical reality: Albanian remains a low-resource language in the NLP landscape. Our evaluation with Word Error Rate (WER), acoustic fidelity, and prosodic alignment shows significant gaps in both synthesis and recognition. These findings underscore the urgent need for targeted research, datasets curation, and model development for Albanian, one of the oldest and most culturally rich languages in the world. We’ll be presenting our findings at the READ International Scientific Conference 2025, hosted by the University of Tirana and Albanian - American Development Foundation (AADF). If you're passionate about linguistic diversity, low-resource language modelling, or the future of inclusive AI, we’d love to connect. Special call for Albanian researchers working in this area, lets join forced to make Albania a resourceful language! Hadeel Saadany Eglantina Gishti Kreshnik Vukatana Enrik Keco Artid Boonrerng Kejvi X.

Proud to be part of this research on Albanian as a low-resource language, exploring Text-to-Speech and Speech-to-Text performance. Our results show clear gaps as Albanian needs more data, research, and collaboration to advance in the AI field. These findings will be presented at the READ International Scientific Conference 2025, hosted by the University of Tirana and AADF. #AI #NLP #AlbanianLanguage #Research #SpeechTechnology #Innovation

Some names leave a mark long after the session ends. At last year’s Bootcamp, Ife Adebara, PhD inspired a wave of young innovators with her bold vision for African languages in technology. This year, she returns to #DSNAIBootcamp2025 — not just as a speaker, but as a symbol of continuity, possibility, and progress. Dr. Adebara is the Co-Founder of EqualyzAI and a globally recognized researcher in natural language processing, linguistics, and language policy. Her groundbreaking work has shaped tools like AfroLID and Serengeti, covering 517 African languages and earning global honors, including the IRCAI - International Research Center on Artificial Intelligence under the auspices of UNESCO award for Top 10 Outstanding Global AI Solutions. From presenting at ACL, EMNLP, and UNESCO’s LT4ALL, to being featured on CBC News and Global News Canada, she has shown the world that African voices and languages belong in the future of AI. As she steps onto the Bootcamp stage once again, one thing is certain: the conversation around AI in Africa will never sound the same. #datasciencenigeria #DSNAIBootcamp2025 #AIforALL 

Hyphens or No Hyphens – AI Stigmatization ? Could AI change the way we write and communicate? When using one of the public LLMs for text generation, you may have noticed that they often include a lot of hyphens in their outputs. If we think about it, most LLMs are trained on reputable books and well edited sources, where grammar and punctuation are generally of high quality. Naturally, the models replicate those patterns. Personally, I use LLMs to check the grammar of my texts, emails, or posts (please don’t overuse them for direct messages). I’m not a native English speaker, so my grammar is not perfect. To be honest, even in Spanish, which is my native language, grammar is not my strongest skill, so I use LLMs for that as well. I don’t have anything against hyphens, and in Spanish text generation this is usually not an issue(we don't use it). I actually think they can sometimes add clarity (I support hyphens). Still, I believe their use should be a personal choice of the author. You can always use alternatives like commas or parentheses instead. I found an interesting Reddit thread discussing this topic: https://lnkd.in/dJYgVGvJ (I used this prompt for this text "check the grammar, please don´t use hypens") #NLP #Curiosity #TextGeneration #LLMs #Hyphenation

For those interested in AI, its possibilities and its limitations, this lecturers is for you! Sponsored by the Department of African studies at The Ohio State University, it takes place next Wed (Oct 29) at 12pm EST. More.details, a flyer and a zoom link follows the abstract below: "Going Beyond Text with AI for African Languages" Artificial intelligence (AI) has made great progress in understanding and generating written language, but in Africa, communication is much richer than text. From the rhythm of our speech to the meanings carried in our names, images, and sounds, much of African expression lives beyond the written word. This lecture explores how AI can move beyond text to truly reflect the realities of African life by learning from our voices, our accents, our stories, and our visual world. It highlights efforts to build technologies that can understand Africa's rich cultural identity. By looking at the next generation of language and multimodal technologies, this talk invites us to imagine an Africa-centered future for AI, one where technology not only speaks our languages, but also listens, sees, and understands in our own ways. Abraham Toluwase Owodunni is a PhD student in Computer Science & Engineering at The Ohio State University, where he is advised by Prof. Sachin Kumar. His research focuses on efficient multilingual representation learning and understanding how linguistic knowledge can transfer across models, especially in low-resource and African language settings. Formerly, he worked as a Lead Machine Learning Researcher at Intron, developing language and speech models tailored to African languages and accents, including work spanning the MENA region. Lunch will be provided. Zoom link: https://lnkd.in/ez59kG2j

We are taking a new direction in LLM Research: Language-Informed Models + Layer-wise Semantic Dynamics (LSD) Over the past months, I’ve been working on Layer-wise Semantic Dynamics (LSD) — a framework that treats large language models (LLMs) not as black boxes, but as dynamical systems whose hidden states trace semantic trajectories across layers. What LSD revealed is fascinating: ✅ Factual answers converge smoothly in semantic space. ✅ Hallucinations drift erratically, showing instability. This shift — watching the path rather than only the output — has opened a new lens on LLM interpretability and safety. Now, a new intuition is emerging. What if certain languages naturally produce more efficient, stable trajectories? Arabic, for example, has unique structural properties: ✅ Root-based morphology that clusters meaning. ✅ Semantic density that packs rich information into fewer words. ✅ Flexible syntax that encodes relationships differently than English. If these properties lead to smoother convergence in LSD trajectories, we may be at the start of a new research area: Language-Informed LLMs — models that explicitly leverage the structure of natural languages (like Arabic) to reason more efficiently and faithfully. 👉 This could mean: ✅ Fewer layers needed to reach factual convergence. ✅ More robust embeddings aligned with meaning. ✅ Lower hallucination risk, thanks to built-in semantic efficiency. This is still a hypothesis, but it feels like a promising and exciting direction — one that combines linguistics, deep learning, and interpretability into a framework for safer and more efficient AI. I’d love to hear from fellow researchers, linguists, and ML engineers: ✅ Have you observed structural advantages in morphologically rich languages? ✅ Could language-informed architectures be the next leap beyond data scaling? Let’s explore this together. #AI #LLM #Interpretability #Hallucination #Linguistics #Arabic #DeepLearning #Research