Tethering Broken Themes: Aligning Neural Topic Models with Labels and Authors