Metaphor Understanding Challenge Dataset (MUNCH)
收藏资源简介:
MUNCH数据集由阿姆斯特丹大学创建,旨在评估大型语言模型(LLMs)对隐喻理解的能力。该数据集包含超过10,000个包含隐喻使用的句子的同义改写,以及1,500个包含不恰当同义改写的实例。这些不恰当的同义改写被精心挑选,以作为控制条件,确定模型是否确实进行了完整的隐喻解释,还是仅仅依赖于词汇相似性。所有的恰当和不恰当的同义改写都经过了人工标注。隐喻句子覆盖了4种不同体裁(学术、新闻、小说和对话)的自然隐喻使用,并展示了不同程度的新颖性。该数据集的应用领域包括隐喻理解、语言模型评估和自然语言处理任务,旨在解决隐喻理解的挑战和提高语言模型的性能。
The MUNCH dataset was created by the University of Amsterdam to evaluate the metaphor comprehension capabilities of large language models (LLMs). It contains over 10,000 paraphrases of sentences containing metaphorical usage, as well as 1,500 instances of inappropriate paraphrases. These carefully selected inappropriate paraphrases serve as control conditions to determine whether models truly perform complete metaphorical interpretation, or merely rely on lexical similarity. All appropriate and inappropriate paraphrases have been manually annotated. The metaphorical sentences cover natural metaphorical uses across four distinct genres: academic, journalistic, fictional, and conversational, and exhibit varying degrees of novelty. The application areas of this dataset include metaphor comprehension, language model evaluation, and natural language processing tasks, aiming to address the challenges of metaphor comprehension and improve the performance of language models.

- 1Metaphor Understanding Challenge Dataset for LLMs阿姆斯特丹大学,荷兰 · 2024年



