Close enough isn’t good enough in word learning: successful cross-situational word mappings are semantically independent of previous mappings

Children often encounter new words in referentially and semantically ambiguous environments. Thus, they will generally make many incorrect guesses about a word’s meaning before arriving at its correct meaning. Here, we asked whether these initial incorrect guesses might nevertheless be useful to learners by providing information about a word’s semantic neighborhood (e.g., if most guesses were food items, perhaps the word has a food-related meaning). To test this, we analyzed datasets from previous tasks in which adults guessed the word which caregivers uttered in interactions with their children. We first tested whether adults’ incorrect guesses are, indeed, semantically similar to the correct meaning. In Study 1, we established that learners’ incorrect guesses were semantically similar to the target word. We then asked whether adults successfully used these semantically similar guesses as “stepping-stones” to arrive at the correct meaning across exposures. Study 2 showed that overall, learners’ guesses were semantically consistent across exposures. However, this effect was small, and correct guesses were not judged to be similar to learners’ prior, incorrect guesses. Moreover, Study 3 revealed that semantically close-to-target guesses did not improve learners’ subsequent accuracy. Thus, even adult word learners fail to use semantic similarity in cross-situational word learning. Study 4 confirmed this result in a new word learning experiment: even for maximally similar meaning pairs, adults failed to generate thematically or taxonomically similar meanings across exposures. While learners’ incorrect guesses tend to be similar to the correct meaning, learners do not successfully use this information to learn words across exposures.