Muhammad AbdulMageed

"Which language is this?" sounds like the easiest question in NLP. For most of Africa's languages, it still has no reliable answer. Today/Tuesday at #EMNLP2026, @SangyooonKwon and I will present AFROSCOPE, our framework for identifying African languages at scale. The gap is real. Existing language ID systems cover a fraction of the continent's languages and mix up close relatives, and the data behind them leans heavily on Bible translations. That matters because everything downstream, from translation to speech, starts with knowing what language you are looking at. AFROSCOPE covers 640 languages across 8 families and 7 scripts. We deduplicated and rebalanced the data so that Bible text makes up about half of it rather than three quarters or more, and we checked for leakage: zero overlap with all six evaluation benchmarks, where a widely used resource overlaps by up to 14%. On top of the data sit a family of fine-tuned models and MIRROR, a second pass that pulls apart the languages a base model confuses. The results: 97.9 macro-F1 on our test set against 71.9 for GlotLID-M, and the top score on 5 of 7 evaluation sets. Frontier LLMs manage 47.4 on the same task. The finding I keep coming back to is that script and close relatedness, not language family, are what predict whether one language helps another. Come find Sang and myself at: 📍 Poster Session F, Hall H, 🕐 Today/Tuesday Oct 27, 09:00 to 10:30. Led by Sang Yun Kwon with @arelmadany at #UBC. This work was partially supported by CLEAR Global via the Gates Foundation. Many thanks to Christian Resch and Howard Lakougna for related discussions. 📄 Paper: lnkd.in/gW9QMZpx 💻 Code and models: lnkd.in/gW-4BPMu #NLProc #AfricanNLP #LanguageIdentification #UBC

Today at #EMNLP2026 in Budapest, we present DRFLOW, a deep research benchmark for personalized workflow prediction. Deep research systems are good at writing reports, but many enterprise tasks need something else: an ordered list of steps that fits one specific person. Think, for example, which forms does this employee file, in what order, and which branch of the company process applies to their case? Answering that means digging through policies, emails, and chats, ignoring distractors, and committing to actions rather than prose. This is where DRFLOW comes in. As a benchmark, DRFLOW has 100 tasks across B2B, B2C, education, healthcare, and legal settings, with 1,246 reference steps grounded in more than 3,900 sources and seven diagnostic metrics. A mixed split adds artifacts from other questions, which makes the tasks much harder. We also introduce DRFA, a workflow-oriented reference agent. Averaged over four LLM backbones, it beats a strong deep research baseline by about 10 F1 points and nearly doubles condition resolution, yet the best configuration (GPT-5.2) reaches only 63.9 F1, and scores drop further on the mixed split. Both agents stay highly factual. The gap is in deciding which steps apply to this user. Come by: 📍 Poster Session F, Hall H 🕐 Today (Tuesday, Oct 27), 09:00 to 10:30 Work by our PhD student @tawkat97, with Raymond Li, Laks V.S. Lakshmanan, and Issam Laradji. A ServiceNow AI Research and UBC collaboration. 📄 Paper: lnkd.in/gyTQisBH #NLProc #DeepResearch #LLMAgents #UBC

UBC NLP will have a strong footprint in Budapest, with 21 papers at #EMNLP2026! 🎉 10 Main Conference, 8 Findings, 1 TACL, and 2 workshop papers (NLLP and REALM). The work covers a lot of ground: 🌍 Culture and values: whether LLMs actually apply the cultural context they infer, South Asian meme and humor understanding, human value geometry in steering space, and value drift during post-training 🗣️ Language diversity: dialectal Arabic speech recognition (Bulbul), the linguistic landscape of Africa (AFROSCOPE), and LLM agents as computational typologists 🔎 Retrieval and RAG: temporal reasoning (TEMPO), biomedical constraints (OntologyBench), paragraph-level and long-document retrieval, and RAG on Canadian case law 🤖 Agents and reasoning: financial and numerical reasoning (MoCA-Agent), computer-use agents, deep research agents, inference as tree search, and claim verification over tabular data 📝 And more: output homogeneity in base models, discharge summarization, and large-scale authorship attribution Huge congratulations to all the students, postdocs, and collaborators behind these papers, and to my colleagues across UBC. If you're in Budapest, come find us at the talks and posters. And stay for ArabicNLP 2026, co-located with EMNLP!