Core Concepts
Large Language Models (LLMs) are evaluated for table editing tasks using the WikiTableEdit dataset, showcasing challenges in editing irregular tables.
Abstract
Tabular data is crucial for organizing and analyzing information.
Manual table editing is laborious, especially for irregular tables.
WikiTableEdit dataset introduces natural language instructions for table editing tasks.
Evaluation of LLMs on WikiTableEdit dataset reveals challenges in editing irregular tables.
Proposed metric Table Edit Distance (TED) for evaluating table generation effectiveness.
Contributions include dataset creation, metric design, and LLM evaluation.
Experiments show the need for improvement in LLMs for table editing tasks.
Comparison of model performance on regular and irregular table editing.
Supervised fine-tuning significantly improves model performance.
Related works focus on TableQA and Data2text, not table editing.
WikiTableEdit aims to advance research in table editing tasks.
Stats
WikiTableEdit는 194,996개의 학습 데이터 인스턴스와 28,706개의 테스트 데이터 인스턴스를 포함한다.
GPT-3.5-turbo는 EM에서 11.08의 성능을 보여준다.
LLaMA2-7B는 정규 테이블과 불규칙 테이블 편집 간의 성능 차이가 없다.
Quotes
"Our primary contributions can be summarized as follows: We introduce the task of table editing and construct a high-quality dataset WikiTableEdit to support this endeavor."
"The current results indicate that existing models still have a long way to go in terms of table editing."