All-in Text: Learning Document, Label, and Word Representations Jointly