Open-Domain Question Answering over Tables with Large Language Models

Open-domain question answering (ODQA) over tables has received attention in recent years. This task requires retrieving evidence from huge amounts of semi-structured tabular data to answer natural language questions. Previous open-domain table QA methods typically focus on customized models, which require a significant amount of annotated training data. Furthermore, most existing models struggle to handle different types of complex problems due to limited reasoning ability and flexibility. Some of them attempt to expand the training and pre-training modules, but this approach increases costs and limits practicality. To alleviate the above challenges, we introduce Generating Identifiers and Selecting chunks for Tables (GIST), which leverages the potential of large language models (LLMs) for the open-domain table QA task. With the assistance of designed prompts and a powerful LLM, GIST first generate question identifiers, and process them for table retrieval. Then we select the candidate evidence to reduce interference caused by the vast amount of data and support subsequent reasoning. Finally, the LLM is prompted to extract the correct answer. Experimental results illustrate that our method achieves competitive end-to-end QA performance on two open-domain table QA datasets. In particular, our method significantly outperforms the current state-of-the-art on NQ-TABLES.