👇 Watch the trailer + subscribe so you don’t miss the full episode!

youtu.be/Bh3bT-r4aH8

🎧 Coming soon — don’t miss it!

🎬 YouTube - youtube.com/@WomeninAIRe...
🎙️ Spotify - open.spotify.com/show/51RJNlZ...
🎧 Apple Podcasts - t.co/IZSYvx3YlI

Annie's research focuses on language diversity, multilinguality, low-resource languages and multicultural fairness in AI systems.

📘 Paper: aclanthology.org/2024.emnlp-m...
🎬 YouTube: www.youtube.com/watch?v=RKIv...
🎧 Spotify: open.spotify.com/episode/3IvN...
🍎 Apple Podcasts: podcasts.apple.com/ca/podcast/w...
#WiAIR #AIResearch #VLM #EMNLP2024 #MulticulturalAI #GLOBALRG
(8/8🧵)

🌍 Models perform strongest on North American & European images, and weakest on East Asia, South Asia, Africa & SE Asia — highlighting cultural bias.
(7/8🧵)

📊 Beyond accuracy, diversity@k measures the cultural range in top-k retrieved results.
It reveals cultural gaps that accuracy alone doesn’t show.
(6/8🧵)

🔎 Cultural Visual Grounding Task
Tests whether models identify culture-specific elements (e.g., molinillo, knafeh, aarti thali) within images.
Performance varies across regions.
(5/8🧵)

📸 Universal Concepts Task
Looks at how models retrieve images for concepts like wedding, breakfast, & funeral.
Findings show limited cultural variation, often leaning Western.
(4/8🧵)

🧪 The study introduces GLOBALRG, a benchmark across 50 countries & 10 regions, with two tasks:
🔸 Cross-cultural image retrieval
🔸 Cultural visual grounding (culture-specific elements from 15 countries)
(3/8🧵)

📄 From Local Concepts to Universals: Evaluating the Multicultural Understanding of Vision-Language Models
The paper examines whether VLMs represent universal concepts in culturally diverse ways, rather than defaulting to one dominant view.
(2/8🧵)

🎧 Don’t miss this #WiAIR episode with Dr. Vered Shwartz!
🎬 YouTube: www.youtube.com/watch?v=RKIv...
🎙 Spotify: open.spotify.com/episode/3IvN...
🍎 Apple: podcasts.apple.com/ca/podcast/w...
(6/6🧵)

📊 INFOGAP exposes how Wikipedia’s “neutral point of view” shifts by language and culture, revealing hidden editorial biases in multilingual knowledge. (5/6🧵)

💬 Facts with polarizing connotations (positive or negative) are excluded more often than neutral ones across all language pairs. Russian LGBT biographies disproportionately share negative facts with English versions compared to non-LGBT bios, revealing selective editorial patterns. (4/6🧵)

🧩 Using 2,700 LGBT biographies, the study finds massive asymmetries:
Only 35% of English facts appear in French and 23% in Russian.
English pages contain far more of the others’ content. (3/6🧵)

🔍 INFOGAP combines cross-lingual fact alignment using multilingual embeddings with LLM-based entailment verification to determine what's shared, omitted, or reframed between English, Russian, and French articles. (2/6🧵)
📄 Paper: arxiv.org/abs/2410.04282

🎧 Listen to our conversation with Dr. Shwartz:
YouTube: www.youtube.com/watch?v=RKIv...
Spotify: open.spotify.com/episode/3IvN...
Apple: podcasts.apple.com/ca/podcast/w...
(7/7🧵)

This also connects with our recent Women in AI Research episode with Dr. Vered Shwartz, where we explore how AI systems trained mainly on Western data can shape language, norms, and expectations worldwide. 🌍✨
(6/7🧵)

Responsible AI isn’t only about avoiding harmful outputs — it also requires careful thinking about how AI interacts with people and the role it takes in their lives. 🔍
(5/7🧵)

This influence becomes especially concerning when people turn to AI for emotional support. Strong safeguards and transparency are needed to protect users. ⚠️
(4/7🧵)