The Growth of Children’s Semantic and Phonological Networks: Insight From 10 Languages

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

Integrative Statement

What factors shape vocabulary learning over the course of early childhood? To investigate this question, one research strategy focuses on studying the properties that make particular words easy or hard to learn. This strategy allows the use of large-scale datasets like CHILDES (MacWhinney, 2000) and Wordbank (Frank et al., 2016) to study average patterns of acquisition across children. For example, within a lexical category, words that are more frequent in child-directed speech are acquired earlier (Goodman et al., 2008). Other factors include word length and concreteness (Braginsky et al., 2016). Besides these intrinsic properties, previous work has shown that some properties of the lexical network also influence the age of acquisition. For instance, children tend to produce words that have higher connectivity in the network earlier, both in the semantic and the phonological domains (Engelthaler & Hills, 2017; Hills et al., 2010; Hills et al., 2009; Stella et al., 2017; Storkel, 2009).

While most studies have focused on the static connectivity in the end-state lexical network, a few have investigated the underlying developmental process. In particular, Steyvers & Tenenbaum (2005) suggested that the observed effects of connectivity are the consequence of how the lexical network gets constructed in the child’s mind. According to this explanation, highly connected words in the child’s lexicon tend to “attract” more words over time, in a rich-get-richer scenario (Barabasi & Albert, 1999). We will call this scenario “internally-driven growth.” However, Hills et al. (2009) found that what biases the learning is not the connectivity in the child’s internal lexicon but, rather, external connectivity in the learning environment. We will call this second scenario “externally-driven growth.”

Hills et al. (2009) tested only English-speakers’ semantic networks, using children’s time of production as an index of the age of acquisition. In this study we test the generality of this finding. We study both phonological and semantic networks, we test both production and comprehension data, and finally, we test growth in ten languages. To this end, we followed the research strategy outlined in Hills et al. (2009) and we used data from Wordbank, an open cross-linguistic database of Communicative Development inventories (CDI, Fenson et al., 1994) curated by our lab.

Our findings (Figure 1) replicate the results obtained by Hills et al. (2009) with English semantic networks. Besides, we had three novel findings: 1) Phonological networks also grow primarily via the externally-driven mechanism, 2) The externally-driven mechanism accounts for both the timeline of words’ production and comprehension, and 3) The results generalize quite well to other languages. Moreover, when pitted against other known predictors of age of acquisition (word frequency and length), the effect of network growth shows a cross-linguistic variation, predicting learning in some languages, but not in others (results not shown here). This variation might be due to low statistical power. In fact, when data are pooled across languages (with language as a random effect), both phonological and semantic networks predicted learning (Figure 2).

Authors

• Abdellah Fourtassi, Stanford University
  Presenting Author

• Yuan Bian, Massachusetts Institute of Technology
  Non-Presenting Author

• Michael C. Frank, Stanford University
  Non-Presenting Author