Knowledge Unit (KU) is the smallest integral knowledge object in a given domain. Knowledge unit relation recognition is to discover implicit relations among KUs, which is a crucial problem in information extraction. This paper proposes a knowledge unit relation recognition framework based on Markov logic networks, which combines probabilistic graphical models and first-order logic by attaching a weight to each first-order formula. The framework is composed principally of structure learning, artificial add or delete formulas, weight learning and inferring. According to the semantic analysis of KUs and their relations, ground predicate set is first extracted.