Machine Learning Algorithm Identifies Tweets Sent Under the Influence of Alcohol

Interesting article posted recently in MIT Technology Reviews. What kind of metrics would help detect such tweets? Which algorithm would you use? I am thinking about tweets indexation, The same NLP (natural language processing) technique can be applied to email messages and other texts produced by users, maybe even to detect if a piece of code was written when the programmer was drunk.This would require the use of a training set, to train the algorithm. But no matter the ML algorithm used, you will need to work with a training set anyway.