Gaussian Embedding

A Region-Based Representation (Vilnis & McCallum, ICLR 2015, Word Representations via Gaussian Embedding) that represents each word as a Gaussian distribution rather than a point. The mean places the word in space; the variance encodes the spread / specificity of its meaning — semantically narrow words get low-variance Gaussians, broad words get high-variance ones. This lets the representation express uncertainty and (asymmetric) containment that a point vector cannot.

It predates Box Embedding as a way to capture hierarchy and overlap. The trade-off versus boxes: computing overlap between Gaussians is more expensive than the per-dimension min/max intersection of axis-aligned boxes, which is a major reason box methods became more widely applied.

Related Concepts

• Region-Based Representation — the family this belongs to
• Box Embedding — box-shaped alternative with cheaper overlap
• Poincaré Embedding — hyperbolic alternative for hierarchy
• Embeddings — point-vector baseline

Articles

• Express Words in a Box - Understanding Box Embedding from the Basics — Shun Tsukagoshi; introduces Gaussian Embedding as a precursor to boxes

People

• Luke Vilnis — lead author (also Box Lattice)