NeurIPS 2020

BoxE: A Box Embedding Model for Knowledge Base Completion

Meta Review

The paper aims to improve knowledge base modelling. In this regards, authors propose a rather ingenious use of box embeddings as the latent representation for the relations. Specifically, each n-ary relation is represented by n boxes and each entity is represented by two vectors. Having a pair of vectors is very powerful, as they allow us to model complex interactions across entities. In particular authors show how their proposed box embeddings can simultaneously handle symmetry, asymmetry, anti-symmetry, and transitivity. No previous framework is claimed to be as flexible nor capable of handling all these patterns. The reviewers also liked the thorough study of theoretical properties. Paper also suggests ways to incorporate prior knowledge into the proposed box embeddings. Finally experiments for link prediction and rule injections were carried out on standard KB datasets. On high degree relations proposed method improves state-of-the-art results significantly as expected from design and injecting prior knowledge as rules also seem to help significantly. Overall reviewers reached a consensus to accept the paper. For the camera ready version, please tune baselines more for Yago and also mention that results are your evaluations and not as reported in original paper (as is the case for other datasets). For example I could get better result for Yago using RotatE.