摘要：Probabilistic models that associate annotations to sequential data are widely used in computational biology and a range of other applications. Models integrating with logic programs provide, furthermore, for sophistication and generality, at the cost of potentially very high computational complexity. A methodology is proposed for modularization of such models into sub-models, each representing a particular interpretation of the input data to be analysed. Their composition forms, in a natural way, a Bayesian network, and we show how standard methods for prediction and training can be adapted for such composite models in an iterative way, obtaining reasonable complexity results. Our methodology can be implemented using the probabilistic-logic PRISM system, developed by Sato et al, in a way that allows for practical applications.
