A number of modern word learning theories posit statistical processes in which knowledge is accumulated across many exposures to a word and its potential referents. Accordingly, words do not go directly from unknown to known, but rather pass through intermediate stages of partial knowledge. This work presents empirical evidence for the existence of such partial knowledge, and further demonstrates its active driving role in cross-situational word learning. Subsequently, an incremental model which leverages its partial knowledge of word-object mappings from trial to trial is shown to account well for the data. In contrast, models which do not do so cannot explain the data. These results confirm crucial assumptions made by statistical word learning models and shed light on the representations underlying the acquisition of word meanings.