Enhancing Interpretability using Human Similarity Judgements to Prune Word Embeddings