Abstract

Financial statements are rich sources of information for market analysis and investment decisions. Most financial statements consist of unstructured text such as business descriptions and structured tables presenting numerical values of financial metrics. Recent studies showed the challenge of answering financial questions due to the difficulty of numerical reasoning over unstructured and structured data. Many novel methods have been introduced to solve this task. However, most existing approaches are data-demanding, which requires a significant amount of annotation effort. We propose a new system that answers questions with case-based reasoning (CBR) to alleviate this issue. CBR is a class of approaches that solve new problems with solutions to existing problems. We propose to leverage CBR such that annotated questions will be retrieved to provide candidate program patterns to an unseen question. The system leverages the program patterns as auxiliary knowledge to generate executable mathematical programs. Our approach decomposes the task into three sub-steps: case selection that provides operation steps for the question, fact retrieval that identifies relevant facts, and program generation that completes the operation steps with the retrieved facts. We conduct low-resource experiments on public financial questionanswering datasets and discuss the usefulness of the system.

Date

March 17, 2026

Authors

Kexuan Sun, Jay Pujara

Conference

1st Workshop on Robust NLP for Finance (RobustFin) @ KDD