KFEX-N : A table-text data question-answering model based on knowledge-fusion encoder and EX-N tree decoder

Abstract

Answering questions about hybrid data combining tables and text is challenging. Recent research has employed encoder-tree decoder frameworks to simulate the reasoning process of arithmetic expressions for generating answers. However, this approach overlooks the inherent diversity of expressions; there might be multiple valid reasoning paths, leading to a decrease in the accuracy of inferred expression trees. Moreover, encoders, lacking rich domain knowledge, struggle to capture deep relationships between questions and supporting evidence; this limitation results in models making errors when selecting operation units. In this paper, we propose a Knowledge-Fusion encoder and EX-N tree decoder table-text data question-answering model(KFEX-N). During the encoding process, the integration of traditional encoders with cross-fusion encoders forms a knowledge-fusion encoder, endowing the model with rich domain knowledge and enhancing its understanding of the operational units required to answer questions. Additionally, we propose an EX-N tree decoder. It reduces the diversity of inference paths through a constrained structure and mitigates the occurrence of final answer errors resulting from decoding errors. We validate our model using publicly available Table-Text QA datasets (TAT-QA and Fin-QA) and achieve state-of-the-art performance.