This paper presents an active learning approach to the problem of systematic noise inference and noise elimination, specifically the inference of Associated Corruption (AC) rules. AC rules are defined to simulate a common noise formation process in real-world data, in which the occurrence of an error on one attribute is dependent on several other attribute values. Our approach consists of two algorithms, Associative Corruption Forward (ACF) and Associative Corruption Backward (ACB). Algorithm ACF is proposed for noise inference, and ACB is designed for noise elimination. The experimental results show that the ACF algorithm can infer the noise formation correctly, and ACB indeed enhances the data quality for supervised learning.