A New Perspective of Statistical Modeling by PRISM

PRISM was born in 1997 as a symbolic statistical modeling language to facilitate modeling complex systems governed by rules and probabilities [Sato and Kameya, 1997]. It was the first programming language with EM learning ability and has been shown to be able to cover popular symbolic statistical models such as Bayesian networks, HMMs (hidden Markov models) and PCFGs (probabilistic context free grammars) [Sato and Kameya, 2001]. Last year, we entirely reimplemented PRISM based on a new tabling mechanism of B-Prolog [Zhou and Sato, 2002]. As a result, we can now deal with much larger data sets and more complex models. In this paper, we focus on this recent development and report two modeling examples in statistical natural language processing. One is a declarative PDCG (probabilistic definite clause grammar) program which simulates top-down parsing. The other is a left-corner parsing program which describes a bottom-up parsing that manipulates a stack. The fact that these rather different types of modeling and their EM learning are uniformly possible through PRISM programming shows the versatility of PRISM.1