Structured learning approaches are able to take into account the relational structure of data, thus promising an enhancement over non-relational approaches. In this paper we explore two document-related tasks in relational domains setting, the annotation of semi-structured documents and the citation deduplication. For both tasks, we report results of comparing relational learning approach namely Markov logic, to non-relational one namely Support Vector Machines (SVM). We discover that increased complexity due to the relational setting is difficult to manage in large scale cases, where non-relational models might perform better. Moreover, our experiments show that in Markov logic, the contribution of its probabilistic component decreases in large scale domains, and it tends to act like First-order logic (FOL).