Abstract: MATH/CHEM/COMP 2002, Dubrovnik, June 24-29, 2002

Modeling distribution of constituent

structure tree for natural language

Bozidar Tepes

Department of Information Science, Faculty of Philosophy, University of Zagreb, I. Lucica 3, HR-10000 Zagreb, Croatia

This paper describes distribution of constituent structure trees of natural language by using Bayesian networks. First part of the paper describes probabilistic context- free grammar (PCFG) together with head-driven phrase structure grammar (HPSG). On these theoretical ideas for modeling formal and natural languages, two Bayesian networks were modeled. First Bayesian network is based on directed acyclic graph (DAG). Second network is based on probabilities from hidden Markov model (HMM) of hidden maximal level structure tree for sentences of natural language. Two networks were tested on Database of grammatical sentences of Croatian language (http://infoz.ffzg.hr/tepes/).