Mostafa Dheghani (UvA) (2015-05-26 13:45 - 14:30)
Hierarchy is an effective and common way of organizing data and representing their relationships in different levels of abstraction. However, hierarchical data dependencies cause difficulty in the estimation of separable models that can distinguish between the entities in the hierarchy. Extracting separable models of hierarchical entities requires us to consider different types of dependencies in the hierarchy and to take their relative position into account.
In this talk, we present generalized language models of hierarchical entities which, inspired by parsimonious language models, provide two-dimensional separable models for hierarchical entities by capturing all, and only, the essential features of their relative position in the hierarchy. This is achieved by iteratively making the models both i) “specific” by adjusting weights of terms already explained by ancestor models, and ii) “general” by adjusting weights of terms already explained by models of descendants.