Compositional Embeddings

Embedding approaches in which the representation of a complex object is built by combining the representations of its parts under well-defined operations, rather than learned as an opaque whole. The goal is compositionality: if you know the embedding of “red” and of “car”, you should be able to derive a meaningful representation of “red car”.

Compositionality is hard for plain point dense vectors — averaging or adding vectors is a crude proxy and loses structure. Region representations are a natural fit:

  • Boxes compose via set operations — the intersection of two boxes approximates the conjunction of their concepts, giving a principled “AND”.
  • This makes box/set-theoretic embeddings a concrete substrate for compositional meaning (conjunction, containment).

Note: this concept is broader than the box-embedding article that seeds it here; the link is conceptual — boxes are one well-behaved way to make embeddings compositional.

Articles