We propose a procedure for sampling representative phrases from any large corpus so that text input researchers can curate their own stimuli for tasks, domains and languages they wish to target using publicly accessible resources. The procedure is based on grounding the notion of representativeness in terms of information theory. Here, you can read the paper and download the code and data.