Can Lexico-semantic Structure in Productive Vocabulary at 18 Months Predict Language Outcomes at Age 3?

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

Integrative Statement

Productive vocabulary at 18 months is highly variable - some toddlers may say only a couple of words, while others may say a couple hundred. This variability, however, does not seem to forecast future language skills. Attempts to associate variability in early vocabulary size with later language outcomes has not yielded strong predictive sensitivity or specificity (Rescorla, 2013). We ask whether digging deeper into the lexico-semantic structure of early productive vocabularies might better forecast later language outcomes at age 3 than simply viewing vocabulary size alone.

We report data from a longitudinal sample of 80 18- to 36-month-old children who were assessed on productive vocabulary skills at 18 months using the MacArthur-Bates Communicative Development Inventory (MBCDI), and then completed the Level 1 core language scale of Clinical Evaluation of Language Fundamentals –P2 (CELF) at 36-39 months to determine the presence or absence of a language delay. We categorized children into three groups based on CELF standard score (SS): Mild delay (77 < SS <= 85; N=6), Moderate delay (SS <=77, N=4), or No Delay (SS >85, N=70).

We then calculated two indices of lexico-semantic structure in each child’s productive vocabulary using two graph-theoretic metrics of semantic structure: mean degree (DEG) and global cluster coefficient (GCC). Mean degree represents the average number of semantic neighbors for each word in the child’s productive vocabulary, while global-clustering coefficient captures general lexical connectivity, by measuring the proportion of semantically connected triads to all possible triads in the lexicon. Thus, these two measures capture both local-level and global-level semantic connectivity in the child’s lexicon. Semantic connections among words were classified according to recently developed database of semantic feature norms for every noun concept on the MBCDI (Peters, McRae, & Borovsky, in prep).

We ran two discriminant analyses to determine whether productive vocabulary percentile, DEG and GCC could together predict later language disorder status. In the first analysis, we sought to determine whether these three measures could distinguish between children who had moderate, mild and no delay. These variables successfully discriminated among groups (Wilks lambda = .82, F(6,150)=2.57, p<.021); reclassification of cases based on canonical variables was moderately successful: 86.25% of cases were correctly reclassified into their original categories (see Table 1 for classification matrix for this model).

Next, we asked whether the combination GCC, DEG and productive vocabulary measures might distinguish between the clinically-relevant group of children who did and did not end up with moderate language delay using discriminant analyses. Though, due to the small diagnostic group size, this analysis indicated that the linear combination of these variables showed only marginal difference between groups (wilks lambda = .92, p=.082), the canonical variable structure achieved excellent classification performance, with 97.5% of cases correctly classified – including all children with moderate language delay. Sensitivity and specificity was excellent (100% and 97.4%, respectively).

The results suggest that lexico-semantic abilities may shed light on mechanisms underlying early language disorders, and can improve early identification. However, replication and extension is necessary to validate the current findings.

Author

• Arielle Borovsky, Purdue University
  Presenting Author