Matching business process models and their node labels plays an important role for business process management. Many matching algorithms using natural language processing (NLP) techniques exist but do not exploit the opportunities of machine learning though it is generally agreed that a learning approach has great potential in the field of NLP. Therefore, we develop a matching approach based on supervised learning using a language-driven similarity function in order to reproduce a human judgement. Additionally, we implement and evaluate our approach using established quality measures, consisting of precision, recall and F-measure. We conduct an evaluation based on real world process models that demonstrates the potential and the limitations of our machine learning approach. The results show a significant learning effect for matching unknown models without predefined rules. The matching quality is comparable to existing matchers. However, the matching quality seems to depend on the one hand on the available training data and on the other hand on the complexity of the chosen similarity function. Further research efforts have to be undertaken in order to improve our approach. This will include developing a more elaborate similarity function containing more linguistic charac-teristics of a label as well as integrating contextual information.