Structured Output Learning with Abstention: Application to Accurate Opinion Prediction