When Interpretability Meets Noise: An LLM-Assisted Hybrid Deep Logical Rule Learning Framework

Abstract

Despite the impressive performance of deep neural networks (DNNs) in natural language processing tasks, their black-box nature greatly limits their applications in life-critical domains. Recently, a growing body of explainable models has been proposed to elucidate the internal working mechanisms of DNNs and enhance their transparency for human comprehension. However, these models fail to provide sufficient insights into model robustness, particularly in learning with noisy labels, thereby limiting their practical utility. To address this gap, we propose a novel LLM-assisted Hybrid Deep logical Rule Learning (HDRL) framework to empower high interpretability with strong noise robustness in high-stakes applications. HDRL adopts a two-phase rule learning paradigm that integrates neural learning with first-order logic, enabling direct interpretability, domain knowledge utilization, and label noise learning. In the first phase, we introduce the Template-based Heuristic Rule Learning (THRL) system, which leverages the predefined rule template and the assistance of large language models (LLMs) to refine and denoise the input space. Based on this, we propose an active rule learning (ARL) system in the second phase to drive logical explanations for model predictions. Extensive experiments across diverse noisy scenarios demonstrate that HDRL can achieve superior performance, high interpretability, and strong noise robustness. This advancement pushes the boundaries of existing rule-based learning methods in noise-prone settings. Furthermore, a case study on the real-world dataset illustrates how HDRL produces accurate and faithful logical explanations for its predictions. Additional counterfactual reasoning on HDRL confirms that these explanations are causally consistent with the underlying decision logic of DNNs.