This research tests the hypothesis that 3- and 4-year-olds can use characteristics of a social context created by adults to learn new words. One of the strategies that a child can use in multi-party conversations is to decide to whom a message (and a new word) is addressed. The ability to do so may simplify word learning situations by making the learning selective and by reducing the amount of perceived words. In the current experiment we test children's ability to learn a new word from a natural conversation when the communicative context is kept constant and when it is altered by adding a new game partner. We predicted that children will differentially interpret verbal messages containing a new word as addressed to them or to the new person, and this will affect their ability to remember the new word. Children heard a new word in one of two conditions: when a communicative context shared with an adult was kept constant and when it has changed (a new adult joined the conversation). We found that 3-year-olds could learn new words only when the communicative context was constant, but 4-year-olds could learn new words in both conditions. A control condition revealed that these findings cannot be explained by task difficulty.