Automated lexical simplification has been performed so far focusing only on the replacement of single tokens with single tokens, and this choice has affected both the development of systems and the creation of benchmarks. In this paper, we argue that lexical simplification in real settings should deal both with single and multi-token terms, and present a benchmark created for the task. Besides, we describe how a freely available system can be tuned to cover also the simplification of phrases, and perform an evaluation comparing different experimental settings.