CodeUnify: A Flexible Code-Driven Framework for Reasoning over Multiple Structured Knowledge Sources

ACL ARR 2025 May Submission7390 Authors

Abstract: Unified structured data question answering task aims to utilize a unified model to answer natural language questions based on different types of structured data. Existing unified structured data question answering methods usually rely on predefined functionalities, which limits their ability to perform complex reasoning beyond these predefined operations. To overcome this limitation, we propose a flexible code-driven framework CodeUnify, which comprises two core modules: CodeSTEP and CRAFT. The CodeSTEP module is a paradigm that generates a complete executable Python code sequences containing a series of step-by-step code-based reasoning query operations based on the question, and CRAFT module (Code-based Reasoning for Adaptive Function Tailoring) can dynamically generate custom code functions for operations beyond the predefined function set, significantly enhancing the flexibility and capability in handling complex reasoning. Comprehensive empirical experiments on multiple structured datasets demonstrate that CodeUnify exhibits superior flexibility and remarkable improvements in complex reasoning scenarios compared to existing unified methods.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: knowledge base QA, reasoning, table QA

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7390