Estimating Subjective Crowd-Evaluations as an Additional Objective to Improve Natural Language Generation.