The modern environment of mobile, pervasive, evolving services presents a great challenge to traditional solutions for enabling interoperability. Automated solutions appear to be the only way to achieve interoperability with the needed level of flexibility and scalability. While necessary, the techniques used to determine compatibility, as a precursor to interaction, come at a substantial computational cost, especially when checks are performed between systems in unrelated domains. To overcome this, we apply machine learning to extract high-level functionality information through text categorisation of a system's interface description. This categorisation allows us to restrict the scope of compatibility checks, giving an overall performance gain when conducting matchmaking between systems. We have evaluated our approach on a corpus of web service descriptions, where even with moderate categorisation accuracy, a substantial performance benefit can be found. This in turn improves the applicability of our overall approach for achieving interoperability in the Connect project.