Yeah, as I said, this was primitive: Stanford NLP was used to extract keywords out of tweets and then it was down to measure the velocity of keywords.