Computational work has suggested that one could, in principle, learn aspects of word meaning simply from patterns of co-occurrence between words. The extent to which humans can do this distributional learning is an open question – artificial language learning experiments have yielded mixed results, prompting suggestions that distributional cues must be correlated with other cues, such as phonological regularities, for successful learning. We conducted a large-scale experiment comparing how distributional learning is afforded by two different types of cues – phonological regularities and semantic coherence. We found that semantic coherence more strongly facilitates distributional learning than onset-consonant phonological regularities.