We propose an approach for modularizing P-log programs and corresponding compositional semantics based on conditional prob- ability measures. We do so by resorting to Oikarinen and Janhunen’s definition of a logic program module and extending it to P-log by introducing the notions of input random attributes and output literals. For answering to P-log queries our method does not imply calculating all the stable models (possible worlds) of a given program, and previous calculations can be reused. Our proposal also handles probabilistic evidence by conditioning (observations).