I'm working at a project that would like to do topic extraction / classification from data, consisting of various NL sources (tweets, social network updates, pastebin). Data are very diverse in content and lenght, and extremely noisy.
The aim is to predict:
a) pertinence to a given set of categories b) distinguish between relevant / non relevant documents for a given application
I'm not really an expert in ML, coming from a more traditional statistics backgroud. Are there techniques that one could use to get some results in term of prediction, lacking a labeled dataset -- which would be very costly to implement right now?