Poster #39 - How to Represent the Toddler Lexicon to Predict Word Learning

• In Event: Poster Session 12

Abstract

Recent work has attempted to understand early language development by representing a child’s lexicon as a network. However, much of the past work modeling children’s vocabulary has utilized adult-based measures. The present investigation asks how the tools and methods used to represent language knowledge impacts this representativeness. We utilize a toddler-directed text corpus we recently created that incorporates child-directed conversations, storybooks for preschoolers, and G-rated movie dialog to approximate the language input an American preschooler might hear. In order to evaluate this corpus in light of vocabulary development, we created semantic networks of words from the MacArthur-Bates Communicative Development Inventory (CDI), a vocabulary checklist. The structure of words and connections in the network were derived using embeddings gathered with a skip-gram neural network (Word2Vec) trained on this Toddler corpus. We then created a second semantic network with the same words, but the connections among words were instead determined using well-known Word2Vec embeddings from a GoogleNews corpus.
To quantify typical growth, we used CDI norms from 16-30 months. Each month, the network consisted of words known according to norms, and any unknown words were available to-be-learned. Using multiple centrality measures, including weighted degree, clustering coefficient, and PageRank, we predicted words-to-be-learned based on which words were considered most central by these different measures. Words ranked as more central in the network may be more similar to words already known, and may have stronger connections for learning and retrieval by the child.
In addition to differences between the two networks (Toddler and Adult), we also investigated two different methods for predicting to-be-learned words. The first, Static, adds words based on the words’ centrality rankings from the full network (“ball” has the same centrality score no matter what words are currently known). The second, Dynamic, adds each unknown word to the current network separately, calculates each associated centrality score, and then chooses the top ranked words as those to-be-learned (“ball” has a different score based on current vocabulary composition). Accuracy was determined by comparing how many words the centrality measures predicted overlapped with words learned according to norms.
Directly comparing the networks and methods to each other, we find that Dynamic predictions perform better than Static predictions, though this is mostly led by a large difference between the Adult predictions, whereas the Toddler Static and Dynamic predictions are comparable. Overall, the Toddler and Adult predictions are no different when compared directly. Comparing predictions to random chance however, we find that the Toddler Static and Toddler Dynamic predictions perform significantly better than random prediction, while the Adult Dynamic only marginally perform better than random, and the Adult Static predictions are no better than random prediction. See Figure 1.
In sum, these results suggest the corpus selection, or how representative the tools utilized are of the population of interest, can significantly impact subsequent model accuracy. Further, we also find that more dynamic ways of quantifying word knowledge result in better models than static measures, though discussion will touch on the large computations needed for dynamic predictions.

Authors

• Jennifer Weber, University of Colorado Boulder
  Presenting Author

• Eliana Colunga, University of Colorado Boulder