First language learners can track word-referent co-occurrence information across situations and evaluate co-occurrence probabilities across situations to determine the best-of-fit mappings. However, cross-situational word learning can be difficult to foreign language learners, because in addition to aggregating information across situations, they have to build robust representations for foreign-sounding words. In Experiment 1, third-grade Mandarin-speaking children learned four English word-referent pairs in two conditions, varying in within-trial ambiguity. In one condition, information about word-referent association was determinable across trials but not within a trial. In the other, word-referent association within a trial was inferable. In Experiment 2, participants learned words in a condition where referential ambiguity across trials, though not within a trial, was reduced by successive presentation of certain word-referent pairs. The results revealed that participants learned more word-referent pairs than expected by chance. Reducing ambiguity within a trial facilitated word mapping and word retention (Experiment 1), but reducing ambiguity across trials did not (Experiment 2). Across two experiments, word mapping and retention performance was associated with phonological awareness, whether measured in children's first language or a foreign language, but not with digit span, suggesting that success in cross-situational word learning in a foreign language rests on specification of word forms.