This paper shows the necessity of distinguishing different referential uses of NPs in Machine Translation. The authors propose a three-way distinction between the generic, referential and ascriptive uses of noun phrases (NPs), and argue that this is the minimum necessary to generate articles and number correctly when translating from Japanese to English. A detailed algorithm is proposed for determining the referentiality of Japanese NPs, based on a defeasible hierarchy of pragmatic rules that are applied top-down, from the clause to the NP. They also sketch the process of generating determiners and number using rules based on the different NP referentialities for a Japanese–English MT system. Using the proposed heuristics has raised the percentage of NPs generated with correct use of articles and number in the Japanese–English MT system ALT-J/E from 65% to 85%.