A small suggestion: I think the way you're describing this as "context" might be misleading. You'll probably want to try the same thing on the output of a BiLSTM or CNN model instead of just the word vectors. I think this is more like an attention mechanism --- like tf-idf, you're learning to downweight words which are less informative. Your method is also very similar in effect to the LexRank line of work, which is more computationally expensive and complicated.