Individual Submission Summary
Share...

Direct link:

Poster #107 - Software for the assessment of language development based on the structure of lexical networks

Thu, March 21, 12:30 to 1:45pm, Baltimore Convention Center, Floor: Level 1, Exhibit Hall B

Integrative Statement

In early language assessments, vocabulary size is generally employed as an indicator for language development in infants. One well-established way to characterize toddlers’ lexicons is to use parent-reported measures that evaluate children’s vocabulary and other communicative skills up to 30 months of age. Perhaps a complementary perspective to examine early language skills is to view language as a self-organized mental lexicon, which can be described by a web-like structure of interacting lexical items. Previous studies have examined language development from this perspective, Beckage, Smith and Hills (2011) provided a network representation of the mental lexicon using the connectivity within the vocabulary of young learners and showed that there are differences in the structure of the vocabularies of children at risk for language impairments and typically developing children. However, it is important that this kind of studies could benefit from enabled instruments to be used by researchers, linguists, teachers, (not only computational engineers or software specialists).
In this work, we present a software for the structural analysis and evaluation of language development (LEXNET-UAEM 1.0.6). To test its functioning and validity, a database of semantic relatedness between words of a specific corpus (CHILDES) was generated. , according to the co-occurrence statistics in a normative language-learning environment consisting of 462,326 words.
Subsequently, 20 children’s vocabularies were collected via a widely used parent checklist, (the CDI-McArthur-Bates). Networks were constructed according to the words in each child’s vocabulary and also in agreement to the co-occurrence statistics of the words in a normative language-learning environment. Each child’s semantic network was derived from the list of words that parents reported their child to use in everyday speech. A random acquisition network for each child was also generated. This network contained a randomly selected set of n words (where n is equal to a given child’s vocabulary size) selected from 258 possible words from the parents checklist and thus the possible words in any child’s network. The random networks were constructed according to the structure of the learning environment. Comparing these networks with each child’s individual lexical network provides a direct measure of the structure differences in two types of networks.
For each network generated (child’s and random networks), three network statistics (in-degree, clustering coefficient, and geodesic distance) were computed in an open-source software for graph and network analysis (GEPHI 0.9.1).
Results showed that the average of clustering coefficient (M=0.085; SD=.048) and in-degree (M=2.753; SD=1.892) of individual lexical networks were significantly higher than those from the corresponding random networks: clustering coefficient (M=.029; SD=.031) and indegree (M=1.110; SD=.979) (p=.006 and p=.025; respectively). Thus children’s networks show more connectivity (in-degree) and more local structure (clustering coefficient) than random networks. This confirms the hypothesis that networks configured with children’s vocabularies show a small world effect (high clustering coefficient and high in-degree). Furthermore, it has been suggested that this small-world structure is characteristic of human language and may be cognitively beneficial to language learning and usage (Cancho & Solé 2001).

Authors