Discovering dependencies in a dataset is at the heart of many data profiling and cleansing efforts. Among the most important dependencies for relational databases are functional dependencies (FDs), which represent constraints between the attributes of a relational data model. FDs can be relaxed on the comparison method and/or on the satisfaction constraint, this relaxation makes their discovery more complex than that of exact dependencies, which is already difficult. In this paper, we propose a methodology to build from a corpus of web tables, a knowledge base that takes the form of a graph whose nodes are semantic types and whose edges represent the existence of a relaxed functional dependency