核心概念
Large Language Models (LLMs) are evaluated for table editing tasks using the WikiTableEdit dataset, encompassing regular and irregular tables.
要約
WikiTableEdit introduces a dataset for table editing tasks, focusing on both regular and irregular tables. The dataset includes various operations such as adding, removing, swapping, reordering, merging, and splitting cells. Large Language Models (LLMs) are assessed on this dataset to demonstrate the challenges of table editing tasks. The study aims to improve models' capabilities in handling diverse forms of tabular data.
統計
WikiSQL dataset consists of 26,531 tables.
WikiTableEdit dataset includes 194,996 training instances and 28,706 testing instances.
GPT-3.5-turbo achieves an EM score of 11.08% on irregular table editing under zero-shot conditions.
引用
"Existing research mainly focuses on regular-shaped tables...editing tables with irregular structures poses a challenge when using code."
"We introduce the task of table editing and construct a high-quality dataset WikiTableEdit to support this endeavor."
"Our primary contributions include introducing the task of table editing and designing a metric Table Edit Distance (TED) for evaluation."