Observability in Hybrid Multi Agent Recurrent Nets for Natural Language Processing

Reading a sequence of lexical items in a sentence is equivalent to providing progressively more data at the input to a Kalman observer. The observer architecture includes a model of the lexical/syntactical sequence generator, with state and output variables, driven by the error between the observed sequence and its evolving 'mirror' within the observer. The theoretical foundations for this observer are put forward and the conditions for observability and controllability of hybrid recurrent nets are explained. Knowledge mining architectures are proposed which consist of an extensible recurrent hybrid net hierarchy of multi-agents where the composite behaviour of agents at any one level is determined by those of the level immediately below.