Cloze Test

Cloze tests, where missing words in sentences must be filled, are increasingly used to evaluate language models' comprehension and knowledge. Current research focuses on improving cloze test generation methods, particularly using transformer-based neural networks and incorporating external knowledge sources like WordNets or knowledge graphs to create more challenging and nuanced tests. This work addresses biases in existing benchmarks, such as the base-rate effect, and explores different test designs, including open-ended cloze tasks and multimodal extensions incorporating images. These advancements refine the evaluation of language models, leading to better understanding of their capabilities and limitations across various tasks.