A Hybrid Search for Complex Table Question Answering in Securities Report

The recent study on a hybrid search method for Table Question Answering (TQA) highlights the growing challenges faced by Large Language Models (LLMs) in accurately interpreting complex table structures. Traditional methods often falter when entire tables are input as text, leading to incorrect answers. The proposed method leverages a combination of a language model and TF-IDF to estimate table headers and select answers based on relevant intersections of rows and columns. Evaluated using the TQA dataset from the U4 shared task at NTCIR-18, this approach achieved an impressive accuracy of 74.6%, significantly outperforming existing models like GPT-4o mini, which only managed 63.9%. This advancement is crucial as it not only enhances the accuracy of data extraction from complex tables but also sets the stage for future improvements in LLMs, potentially incorporating more efficient text-search models to further boost performance.