Stacked generalization learning for document annotation

A method for learning functions which are to be used in annotating a document comprising:

• for a plurality of data elements in a training set of input data, the plurality of data elements being associated with a set of static features, each of the data elements being assigned to one of a plurality of data element types:
• deriving a dependency network in which pairs of the data elements in the training set are linked by dynamic features selected from a set of dynamic features, each of the dynamic features defining a link between a label of a first data element in a pair of the data elements and a label of a second data element in the pair of data elements;
• employing a learning algorithm to return at least one function for each of a plurality of levels, each function expressing a conditional dependency of the labeling of a data element with information from a neighbor data element upon static features of the element;
• wherein at each level after the first level, links between more distant data elements are considered in returning the functions.

A document annotation method comprising:

• training a static model level respective to static features of document elements, the trained static model level configured to assign labels to document elements;
• assembling a stack of model levels having the static model level as the lowest level and further having one or more higher dynamic model levels generated by one or more iterations of a stacked generalization method comprising (i) constructing a set of dynamic features that link elements of a lower level of the stack and (ii) training a dynamic model level respective to the constructed set of dynamic features, the trained dynamic model level configured to assign labels to the document elements;
• iteratively applying the stack of model levels to a document starting with the lowest static model level to infer a collective probabilistic model for the document expressing a joint conditional probability of labeling data elements with labels, each label being derived from a label of a neighbor data element linked to the data element by a dynamic feature in a relational dependency network;
• and
• annotating the document based on the collective probabilistic model.