Efficient Non-parametric Estimation of Multiple Embeddings per Word in Vector Space