Activities at EMNLP & ArabicNLP 2025 -- Research, Benchmarks, and Community Building

EMNLP (Main & Findings)

At EMNLP, our work centered on explainable detection of propaganda and hateful content in multimodal settings:

• MemeIntel: Explainable Detection of Propagandistic and Hateful Memes
  We introduce an explainable framework for detecting propaganda and hate in memes by jointly reasoning over text and image content.
  🔗 Data: https://huggingface.co/datasets/QCRI/MemeXplain
  🔗 Model: https://huggingface.co/QCRI/MemeIntel
  🔗 Paper: https://aclanthology.org/2025.emnlp-main.1539.pdf

• PropXplain: Can LLMs Enable Explainable Propaganda Detection?
  This work explores how LLMs can provide human-interpretable explanations for propaganda detection, moving beyond simple labels toward richer narrative justifications.
  🔗 Data: https://huggingface.co/datasets/QCRI/PropXplain
  🔗 Paper: https://aclanthology.org/2025.findings-emnlp.1296.pdf

ArabicNLP 2025 (Conference & Shared Tasks)

At ArabicNLP 2025, we contributed to both the main conference and multiple shared tasks, with a strong focus on Arabic LLM benchmarking, culture, and multimodality:

• BALSAM: A Platform for Benchmarking Arabic Large Language Models
  BALSAM is a unified platform for evaluating Arabic LLMs across tasks, domains, and varieties of Arabic.
  🔗 Platform: https://benchmarks.ksaa.gov.sa/b/balsam
  🔗 Paper: https://aclanthology.org/2025.arabicnlp-main.21.pdf

• MarsadLab @ AraGenEval (Authorship Style Transfer & Identification)
  MarsadLab systems tackle authorship style transfer and identification, shedding light on stylistic signals in Arabic text.
  🔗 Paper: https://aclanthology.org/2025.arabicnlp-sharedtasks.15.pdf

• MAHED Shared Task: Multimodal Detection of Hope and Hate Emotions in Arabic Content
  We introduce MAHED for multimodal detection of hope and hate, bridging emotion recognition with socially impactful applications.
  🔗 Paper: https://aclanthology.org/2025.arabicnlp-sharedtasks.75.pdf

• CultranAI at PalmX 2025: Data Augmentation for Cultural Knowledge Representation
  CultranAI explores culturally aware data augmentation strategies to better ground LLMs in Arabic cultural contexts.
  🔗 Paper: https://aclanthology.org/2025.arabicnlp-sharedtasks.111.pdf

• MarsadLab at PalmX Shared Task: An LLM Benchmark for Arabic Culture and Islamic Civilization
  We contribute a benchmark focused on Arabic culture and Islamic civilization, designed for evaluating cultural competence in LLMs.
  🔗 Paper: https://aclanthology.org/2025.arabicnlp-sharedtasks.112.pdf

• AYA at PalmX 2025: Modeling Cultural and Islamic Knowledge in LLMs
  AYA investigates how LLMs can represent and reason about cultural and Islamic knowledge in a more faithful and responsible way.
  🔗 Paper: https://aclanthology.org/2025.arabicnlp-sharedtasks.75.pdf

Community & Service

Beyond papers and systems, we have also organized a BoF session:

• Organized BoF: Multilingual & Multimodal LLMs for Everyday Knowledge
  We organized a Birds of a Feather session that brought together researchers working on multilingual and multimodal LLMs, with a special focus on everyday, culturally grounded knowledge. The session highlighted open challenges, shared resources, and opportunities for collaboration across speech, vision, and text.