Reviews: Learning semantic similarity in a continuous space

The paper addresses the problem of learning the representation of questions/sentences as an input unit. Like word2gauss, the embedding is into distributions (as opposed to a single point of a vector space) and in particular into Gaussian distributions. Embedding into a distribution enables the use of a transport distance, such as W2 for computing discrepancy between distributions which has advantages over alternatives such as KL divergence but is expensive to compute. Here the specific constraints on the distributions made the computation of the transport distance efficient. The application considered is question de-duplication.