Using natural language processing to explore 'Dry January' posts on Twitter: a longitudinal infodemiology study