On the Relation between Quality-Diversity Evaluation and Distribution-Fitting Goal in Text Generation