Traversing Knowledge Graph in Vector Space
Explanation of Research paper -> https://arxiv.org/pdf/1506.01094.pdf
- What is Compositional Knowledge Base Representation?
Compositionalization is the approach of using Vector Space Models to complete the Knowledge Base. A Base Vector Space model is considered as a Soft edge Traversal Operator through the Knowledge graph. Compositionalzation involves using base Vector Space models to traverse through the Knowledge Graph and predict the path for each and every path query.
Compositional Knowledge Base tries to reduce the Cascading errors by moving the entities close to the centroid; in other words, compositionalization of Knowledge Graph brings about structural regularization which helps avoiding the Cascading errors as we traverse from one entity to another in the Knowledge graph.
Let us consider Bilinear model for Knowledge Base Completion using Compositionalization. The path query is defined as q = (s/r) where s= entity, r=edge/relation. In order to determine if an entity “t” is connected to entity, “s”, the scoring is performed using the formula,
score(q,t) = (xs^T *Wr*xt)
In the above formula,
xs, xt = Vector representation of entities.
