Does a word-vector dataset that "looks clean" perform better on real applications, justifying the extra attention required for more preprocessing/cleanup?