Matching tables against Knowledge Graphs is a crucial task in many applications. A widely adopted solution to improve the precision of matching algorithms is to refine the set of candidate entities by their type in the Knowledge Graph. However, it is not rare that a type is missing for a given entity. In this paper, we propose a methodology to improve the refinement phase of matching algorithms based on type prediction and soft constraints. We apply our methodology to state-of-the-art algorithms, showing a performance boost on different datasets.