The American Sign Language Knowledge Graph: Infusing ASL Models with Linguistic Knowledge
The American Sign Language Knowledge Graph (ASLKG) enhances the generalizability and explainability of ASL models by incorporating expert linguistic knowledge. Using ASLKG, we achieve 91% accuracy in isolated sign recognition (ISR), 14% accuracy in predicting semantic features of unseen signs, and 36% accuracy in classifying Youtube-ASL video topics.
Jan 22, 2025
Chain-of-Query Prompting for Efficient Small Language Models in Multi-Hop Open-Domain Question Answering
We introduce Chain-of-Query (CoQ), a novel prompting approach that enhances smaller language models (LMs) in multi-hop open-domain question answering (ODQA) by decomposing complex queries into context-aware subqueries, improving efficiency and factual accuracy. CoQ demonstrates up to 5.4% and 11.5% performance gains for small-scale and large-scale LMs, respectively, on benchmark datasets, making it a valuable solution for complex QA tasks.
Dec 15, 2024