Enhancing scientific table understanding with type-guided chain-of-thought

Abstract

Tables in scientific papers convey essential data and insights. Traditional methods struggle with the complexity of modern table data. This study introduces the SciTable-Sowise framework, which utilizes a fine-tuned table classifier to determine the specific type of each table and uses this type information to formulate the Chain-of-Thought (CoT) prompts for large language models (LLMs), significantly enhancing the processing of table content. We constructed the Sci-Table-QA and Sci-Table-Summarization datasets, which comprise 55,000 reasoning QA samples and 5264 summarization samples across multiple disciplines in both Chinese and English. Experimental results show a 7.2 % increase in table reasoning accuracy in Chinese (81.9 %) and a 7.5 % increase in English (83.1 %), surpassing existing models. Our method also enhances summarization performance, as validated by ROUGE, BertScore, and GPT-4o model evaluation metrics (G-Eval-4). This approach demonstrates substantial real-world application potential in scientific research and business analytics, with our datasets publicly available to support future research.