How semantic seeds shape distributional learning: Evidence for semantic generalization over distributional categories

In three experiments, we investigated how a small vocabulary of meaningful words - a semantic seed - shapes distributional learning. Adults were exposed to an artificial language that mimicked Subject-Object-Verb word order and contained pseudowords organized into three distributional categories. Prior to exposure to this language, half of participants were taught the meanings of several Seed words, while the other half were trained only on the form of the same Seed words. Then all participants heard and read sentences from the language, containing both the Seed words and novel words, followed by a task assessing their semantic biases for the novel words. In Experiment 1, we found that learners can use novel words’ distributional histories to make inferences about their meanings, but only when they already had a semantic seed. Experiment 2 demonstrated that the benefit of semantic seeding is not merely the result of familiarity with particular Seed meanings. Experiment 3 showed that semantic seeds influence distributional learning by revealing semantic features of distributional categories, rather than by supporting interpretation of sentences containing semantic seeds. A grammar test in Experiment 3 provides evidence that learners, regardless of whether they have a Semantic Seed, engage in distributional analysis over all sentences. Thus, Semantic Seeds and distributional learning interact to enable learners to discover semantic features of distributional classes, which can be generalized to all words within a distributional class.