Can we train up a sentiment analyzer on the delta between the stories and comments this week vs. all others, and then apply it to discovering "political discussion?"