"The day after the day after tomorrow?" A machine learning approach to adaptive temporal expression generation: training and evaluation with real users