23755432da68528f115c9633c0d7834f-AuthorFeedback.pdf

Reviewer #1: Our key contribution is to explain mathematically how word embeddings of Glove/W2V capture1 semantic properties of words. Whilst some aspects relate to previous empirical observations or hypotheses e.g. GR manifests geometrically37 as a small difference between one PMI vector (or sum of PMI vectors) and another, meaning that the associated38 KL divergence is small and the relevant semantic relationship (similarity, paraphrase or analogy) exists.