A fundamental component of language acquisition involves organizing words into grammatical categories. Previous literature has suggested a number of ways in which this categorization task might be accomplished. Here we ask whether the patterning of the words in a corpus of linguistic input (distributional information) is sufficient, along with a small set of learning biases, to extract these underlying structural categories. In a series of experiments, we show that learners can acquire linguistic form-classes, generalizing from instances of the distributional contexts of individual words in the exposure set to the full range of contexts for all the words in the set. Crucially, we explore how several specific distributional variables enable learners to form a category of lexical items and generalize to novel words, yet also allow for exceptions that maintain lexical specificity. We suggest that learners are sensitive to the contexts of individual words, the overlaps among contexts across words, the non-overlap of contexts (or systematic gaps in information), and the size of the exposure set. We also ask how learners determine the category membership of a new word for which there is very sparse contextual information. We find that, when there are strong category cues and robust category learning of other words, adults readily generalize the distributional properties of the learned category to a new word that shares just one context with the other category members. However, as the distributional cues regarding the category become sparser and contain more consistent gaps, learners show more conservatism in generalizing distributional properties to the novel word. Taken together, these results show that learners are highly systematic in their use of the distributional properties of the input corpus, using them in a principled way to determine when to generalize and when to preserve lexical specificity.

Successful language acquisition hinges on organizing individual words into grammatical categories and learning the relationships between them, but the method by which children accomplish this task has been debated in the literature. One proposal is that learners use the shared distributional contexts in which words appear as a cue to their underlying category structure. Indeed, recent research using artificial languages has demonstrated that learners can acquire grammatical categories from this type of distributional information. However, artificial languages are typically composed of a small number of equally frequent words, while words in natural languages vary widely in frequency, complicating the distributional information needed to determine categorization. In a series of three experiments we demonstrate that distributional learning is preserved in an artificial language composed of words that vary in frequency as they do in natural language, along a Zipfian distribution. Rather than depending on the absolute frequency of words and their contexts, the conditional probabilities that words will occur in certain contexts (given their base frequency) is a better basis for assigning words to categories; and this appears to be the type of statistic that human learners utilize.

We regularly make predictions about future events, even in a world where events occur probabilistically rather than deterministically. Our environment may even be non-stationary such that the probability of an event may change suddenly or from one context to another. 4–6 year olds and adults viewed 3 boxes and guessed the location of a hidden toy. After 80 trials with one set of probabilities assigned to the 3 boxes, the spatial distribution of these probabilities was altered. Adults easily responded to this change, with participants who maximized in the first half (by choosing the most common location at a higher rate than it was presented) being the fastest at making this shift. Only the older children successfully switched to the new location, with younger children either partially switching, perseverating on their original strategy, or failing to learn the first distribution, suggesting a fundamental development in children’s response to changing probabilities.

No abstract

There has been significant recent interest in clarifying how learners use distributional information during language acquisition. Many researchers have suggested that distributional learning mechanisms play a major role during grammatical category acquisition, since linguistic form-classes (like noun and verb) and subclasses (like masculine and feminine grammatical gender) are primarily defined by the ways lexical items are distributed in syntactic contexts. Though recent experimental work has affirmed the importance of distributional information for category acquisition, there has been little evidence that learners can acquire linguistic subclasses based only on distributional cues. Across two artificial grammar-learning experiments, we demonstrate that subclasses can be acquired from distributional cues alone. These results add to a body of work demonstrating rational use of distributional information to acquire complex linguistic structures.