Not All Contexts Are Created Equal: Better Word Representations with Variable Attention