上海科技大学知识管理系统

Large language models (LLMs) have exhibited strong semantic understanding capabilities in interpreting and reasoning for table question answering (TQA). However, they still face challenges with excessively long or complex input tables, particularly when these tables are disorganized or have hierarchical structures. To address these challenges, we propose a new paradigm for TQA, named TableCall, which leverages the tool-using capabilities of LLMs. Specifically, TableCall invokes different tools for different types of table questions, such as SQL, Python, and LLMs, to simplify and enhance the reliability of table understanding. Moreover, to further enhance LLM’s table comprehension capabilities, a few-shot library updating technique is proposed to generate better QA pairs for LLM prompting. Experimental results on both open-domain and specific-domain datasets demonstrate that our approach achieves state-of-the-art performance, outperforming previous methods in terms of accuracy, efficiency, and reliability.