A new paper coauthored by researchers at The University of Melbourne and MBZUAI explores disagreement in human annotation for AI training. The paper treats disagreement as a signal (human label variation or HLV) rather than noise, and proposes new evaluation metrics based on fuzzy set theory. These metrics adapt accuracy and F-score to cases where multiple labels may plausibly apply, aligning model output with the distribution of human judgments. Why it matters: This research addresses a key challenge in NLP by accounting for the inherent ambiguity in human language, potentially leading to more robust and human-aligned AI systems.
This paper explores multilingual satire detection methods in English and Arabic using zero-shot and chain-of-thought (CoT) prompting. It compares the performance of Jais-chat(13B) and LLaMA-2-chat(7B) on distinguishing satire from truthful news. Results show that CoT prompting significantly improves Jais-chat's performance, achieving an F1-score of 80% in English. Why it matters: This demonstrates the potential of Arabic LLMs like Jais to handle nuanced language tasks such as satire detection, which is critical for combating misinformation in the region.
MBZUAI researchers presented a study at NAACL 2024 analyzing errors made by open-source LLMs when solving math word problems. The study, led by Ekaterina Kochmar and KV Aditya Srivatsa, investigates characteristics that make math word problems difficult for machines. Llama2-70B was used to test the ability of LLMs to solve these problems, revealing that LLMs can perform math operations correctly but still give the wrong answer. Why it matters: The research aims to improve AI's ability to understand and solve math word problems, potentially leading to better educational applications and teaching methods.
A new paper from MBZUAI introduces JEEM, a benchmark dataset for evaluating vision-language models on their understanding of images grounded in four Arabic-speaking societies (Jordan, UAE, Egypt, and Morocco) and their ability to use local dialects. The dataset comprises 2,178 images and 10,890 question-answer pairs reflecting everyday life and culturally specific scenes. Evaluation of several Arabic-capable models (Maya, PALO, Peacock, AIN, AyaV) and GPT-4o revealed that while models can generate fluent language, they struggle with genuine understanding, consistency, and relevance, especially when cultural context is important. Why it matters: This research highlights the challenges of building AI systems that can truly understand and interact with diverse cultures, emphasizing the need for culturally grounded datasets and evaluation metrics.
Researchers at MBZUAI have introduced QRAFT, an LLM-based framework designed to automate the generation of fact-checking articles. The system mimics the writing workflow of human fact-checkers, aiming to bridge the gap between automated fact-checking systems and public dissemination. While QRAFT outperforms existing text-generation methods, it still falls short of expert-written articles, highlighting areas for further research.