Learning Rich Hidden Markov Models in Document Analysis: Table Location

hidden Markov models (HMM) are probabilistic graphical models for interdependent classification. In this paper we experiment with different ways of combining the components of an HMM for document analysis applications, in particular for finding tables in text. We show: a) how to integrate different document structure finders into the HMM; b) that transition probabilities should vary along the chain to embed general knowledge axioms of our field, c) some emission energies can be selectively ignored, and d) emission and transition probabilities can be weighed differently. We conclude these changes increase the expressiveness and usability of HMMs in our field.