Figure 2.

Illustration of the ESP architecture to generate 5-dimensional embeddings for a 10-concept vocabulary. In SGNS, an objective is to raise the scalar product between the semantic and context vectors of words that are observed together. With ESP, we aim to raise the $\text{NNHD}—(\mathit{MAX}(0,(1-\frac{2\times \text{hamming}\text{distance}}{\text{dimensionality}})))$— between the semantic vector for the subject of a predication, and the bound product (⊗) of the predicate vector for the predicate concerned and the context vector of the object of this predication. In addition, we aim to raise the NNHD between the context vector of the object of the predication and the product of releasing (⊘) the predicate from the semantic vector of the subject.