A Model-based Vocabulary Structure Intervention for Typically-developing and Late Talkers

• In Event: 2-028 - The Organization of Knowledge: Development and Modeling of Early Semantic Networks

Integrative Statement

Language can be thought of as a highly structured system of relations. Learning a first language, then, becomes a self-propelling process in which initial advantages can further leverage access to learn more complex language. Most children exploit the structure in language, becoming increasingly more efficient language learners, for example by acquiring word-learning biases (Smith, Jones, Landau, Gershkoff-Stowe, & Samuelson, 2002). Late talkers, children below the 20th percentile on productive vocabulary, are not a homogenous group in terms of their developmental outcomes – about half of them go on to catch up to their peers without need for intervention, while the rest may show lasting language delays that eventually impact academic achievement and emotional development among other things (e.g., Desmarais, Meyer, Bairati, & Rouleau, 2008). When learning new words in the lab, late talkers are less efficient than typically developing children, and do not show similar word learning biases (Jones, 2003). Differences in language development may come from differences in children’s ability to exploit this structure or differences in the structure of their early vocabularies. Indeed, the vocabularies of late talkers and typically developing children have been shown to contain different kinds of words (Perry & Samuelson, 2016; Colunga & Sims, 2017) and form different network structures (Beckage, Smith & Hills, 2012). In this longitudinal study, we use individualized network-based models to select target words for children and their parents to learn at home, and observe their vocabulary growth and structure over time.

Forty-six children between 19-20 months of age (including 20 late talkers, below the 20th percentile in productive vocabulary), were assigned to high probability and low probability conditions, depending on whether they were to be given books containing words the models suggested were likely or unlikely to be learned next. The models were used as follows. For each child, we built a semantic network representing their initial vocabulary with the words in their vocabulary (according to parental report) as nodes, and sensory-motor feature similarity (Howell, Jankowicz & Becker, 2005) as the criterion for connecting words (e.g., dog and cat are connected because they share a lot of features, but dog and refrigerator are not because they don’t). Each individual network was grown using a modified preferential acquisition algorithm, in which potential new words that are better connected to known words are more likely to be learned (Beckage et al, 2012). We simulate this process, acquiring the next 50 words, 1000 times, generating an estimated probability that each word will be among the next 50 to be learned, and then selecting words from this list, for each child, depending on the condition they were assigned to, and controlling for word’s age of acquisition. We then created individual story books for each child to take home (see timeline in Fig. 1).

Results (some shown in Fig. 2) suggest that late talkers and typically developing children may benefit from learning different types of words, and specifically that typically developing children may benefit more from learning high probability words than low probability words.

Authors

• Eliana Colunga, University of Colorado-Boulder
  Presenting Author

• Nicole Beckage, University of Wisconsin-Madison
  Non-Presenting Author

• Jennifer Weber, University of Colorado-Boulder
  Non-Presenting Author