Introduction

It is known that children with Autism Spectrum disorder (AS) experience a delay in their vocabulary growth compared to typically developing children (TD). In this study, we investigate if the two population vary only in terms of the time it takes them to learn the same words, or whether they also differ in the way they learn these words.

Previous studies using corss-sectional CDI data have shown that TD children learn words based on their salience in the speech input environment rather than based on their similarity to previously learned words (Hills et al., 2009; Fourtassi et al., 2018).

AS children may learn differently. In fact, there are two reasons they may follow the second strategy (i.e., similarity to known words):

  1. Lower attention to social cues may reduce the influence of salience in the parent’s input
  2. The restrictive and repetitive behavior which manifests, for example, in the insistence on sameness/similarity makes it possible that AS children learn new words that highly resemble known words both semantically and phonologically.

Our analyses show that, while AS children do experience a significant delay with respect to AD children, they follow a similar learning strategy.

Datasets

From NDAR database.

Word and Gestures (WG) dataset

This graph shows the number of words learned as a function of age. Note the sparsity of the sampling: it looks like most data are sampled from only a couple of months.

This graph shows the number of children included in each month. It confirms the sparsity issue: most data are indeed sampled from only two months.

Word and Sentences (WS) dataset

We plot similar graphs for the WS dataset. We still have sparsity, but several months between 0 and 50 are now well represented.

Trajectory of word learning

Here I show the learning trajectories of two words “ball” (a noun) and “play” (a verb). Each point is the proportion of AS children that know the word at a given age. Solid lines represent the logistic regression fit.

For comparison, I also plot (in red) the trajectories for TD children from wordbank.

Note two things:

-Sparsity introduces noise, but the curve is still monotonic as a general trend (although not locally). Compare this to TD curve which is strictly monotonic.

-There is a huge delay in the trajectory of acquisition in AS children: the curve converges to 1 around 100 months of age!

The trajectory of the word “ball”

The trajectory of the word “paly”

The age of acquisition

For each word in the dataset, we define the age of acquisition as the month at which the logistic curve crosses 0.5. Here I plot the density distribution of AoAs in both WG (dotted) and WS (solid) datasets, for both AS and TP children (colors).

One surprising thing about these data is the fact that AS children learn words earlier in WS than they do in the WG dataset (the opposite of the pattern obtained with TD children). That said, keep in mind that the AoA in WG may not be as reliable as WS due to the sparsity issue.

lexical network analysis

Here we used the data to construct semantic and phonological networks following the method outlined in Fourtassi, Bian, and Frank (2018).

Nodes in the network represent words and the edges between the nodes represent semantic or phonological relationship between pairs of words.

For the semantic networks, we used the cue-target relationships in the Florida free association norms. For the phonological relations, we used edit distance computed over the phonological transcription of the words.

We started with exploring the “static” properties of the full, end-state network (constructed using the entire set of words) and then we studied the mechanism of growth.

All analyses involved the concept of “degree.” The degree is defined, for a given word, as the number of other words to which it is related.

First, we sought to replicate with AS children two important findings obtained previously only with TD children: 1) the degree of a word in the end-state network predicts its AoA and 2) the degree distirution of the end-state network follows a power-law. Second, we investigated whether the mechanism of growth is driven more by the structure of previous word knowledege or whether it is influenced by salience in the speech input regardless of the children’s previous knowledge.

Degree vs. Age of acquisition

Here I plot the relationship between the degree z-score and the corresponding AoA for both WG (described as “understands”) and WS (“produces”) in AS children.

We find there to be a correlation between the degree and AoA. The words with the most connections in both semantic and phonological networks tend to be learned earlier.

For comparison, I show below the equivalent plots for TD children.

Degree distribution

Here we examine whether the degree distribution in the semantic and phonological networks follows a power-law. This examination can be crucial to the mechanism of growth. In particular, the absence of a power-law rules out the rich-get-richer growth mechanism (see Fourtassi et al. 2008).

I show a log-log plot of the degree distribution. A power-law should appear as a straight line in such a graph. For comparison, I plot the TD data next to AS data.

Results: AS data are very similar to TD data and, overall, tend to approximate a power-law distribution (after a certain cut-off).

Mechanism of growth

We compare two mechanisms of growth. Given a known lexical network at time t, the word that will be learned at time t+1 can be chosen based on the structure of the existing knowledge (Internal criterion) and/or on the structure of the input (External criterion).

A well-studied example of the internal criterion is based on the rich-get-richer principle: learners select the word that would connect to a word with a high degree in the known (and yet incomplete) network.

An instance of the external criterion is based on salience in the input: learners select a word with a high degree in the input network (constructed from the parents’ speech).

Previous studies (Hills et al., 2009; Fourtassi et al. 2018) explored these mechanisms in TD children and found that, on average, children tend to learn based on the external criterion.

What about AS children?

Models for AS children

For details of how we modeled the Internal and External mechanisms, see Fourtassi et al. (2018).

In brief, we ran a model that is equivalent to a regression in that it fits parameters that characterize some predictors. The set of predictors includes the two learning mechanisms (INT and EXT) used with different information (phonological and semantic). We end-up with 4 predictors: SemINT, SemEXT, phonoINT, and phonoEXT.

The model was trained on the real trajectory of word learning (using AoAs).

If the predictor is different from 0, it means the corresponding mechanism contributes to predicting word learning.

As we see below, AS children follow the external criterion (similar to TD).

Models for TD children

For comparison, I reproduced the results obtained with TD children.

Comparison to other predictors of word learning

We explore if the EXT mechanism predicts word learning when controlling for word frequency and length.

The results, below, show that this mechanism remains predictive (except the for semantics in the WG data).

Here I plot the results of TD children for comparison.