Caregiver linguistic alignment to autistic and typically developing children: A natural language processing approach illuminates the interactive components of language development

Abstract

Background: Language development is a highly interactive activity. However, most research on linguistic environment has focused on the quantity and complexity of linguistic input to children, with current models showing that complexity facilitates language in both typically developing (TD) and autistic children.

Aims: After reviewing existing work on caregiver engagement of children's utterances, we aim to operationalize such engagement with automated measures of linguistic alignment, thereby providing scalable tools to assess caregivers' active reuse of their children's language. By assessing the presence of alignment, its sensitivity to the child's individual differences and how well it predicts language development beyond current models across the two groups, we showcase the usefulness of the approach and provide initial empirical foundations for further conceptual and empirical investigations.

Methods: We measure lexical, syntactic and semantic types of caregiver alignment in a longitudinal corpus involving 32 adult-autistic child and 35 adult-TD child dyads, with children between 2 and 5 years of age. We assess the extent to which caregivers repeat their children's words, syntax, and semantics, and whether these repetitions predict language development beyond more standard predictors.

Results: Caregivers tend to re-use their child's language in a way that is related to the child's individual, primarily linguistic, differences. Caregivers' alignment provides unique information improving our ability to predict future language development in both typical and autistic children.

Conclusions: We provide evidence that language development also relies on interactive conversational processes, previously understudied. We share carefully detailed methods, and open-source scripts so as to systematically extend our approach to new contexts and languages.

Keywords: Autism; Conversation; Language development; Linguistic alignment; Linguistic entrainment; Natural language processing.

Fig. 1.

Example of lexical alignment as cosine similarity. The two utterances are represented as vectors of unique words and their frequency. The angle between the two vectors is calculated (here 45 degrees) and then transformed using the cosine function to scale it between 0 and 1 (here 0.785). The higher the score, the more similar the two utterances.

Fig. 2.

Exact caregiver repetitions in real caregiver-child pairs (red for autism, blue for TD) and 200 randomly sampled surrogate ones (green). The tendency to produce lexical repetitions is displayed in the left panel, syntactic repetitions (excluding lexical ones) in the right panel. Background transparencies display the average rate of repetitions over time for each single real and surrogate caregiver. The x-axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visits 2, 3, 4, 5, and 6.

Fig. 3.

Alignment rate in real caregiver-child pairs (red for ASD, blue for TD) and 200 randomly sampled surrogate ones (green). The tendency to align one’s lexical choices is displayed in the left panel, one’s syntactic structures (adjusting for lexical alignment) in the right panel. Background transparencies display the average rates over time for each single real and surrogate caregiver. The x-axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visit 2, 3, 4, 5, and 6.

Fig. 4.

Panels illustrating the effects of the child’s EL at visit 1 on caregiver lexical alignment rate (left) and syntactic alignment rate (right). The plots extrapolate fictive children from the models, varying only one of the factors at a time (while in real children they tend to co-vary): EL varies while VR and Socialization are kept at population average. The x-axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visit 2, 3, 4, 5, and 6. Note that the y-axes are on different scales to facilitate assessing the effects. See Fig. S11, for analogous plots varying VR and Socialization scores.

Fig. 5.

Alignment level in real caregiver-child pairs (red for ASD, blue for TD) and 200 randomly sampled surrogate ones (green). The tendency to align one’s lexical choices is displayed in the left panel, one’s syntactic structures (adjusting for lexical alignment) in the right panel. Background transparencies display the average levels over time for each single real and surrogate caregiver. The x-axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visit 2, 3, 4, 5, and 6.

Fig. 6.

Panels illustrating the effects of the child’s EL at visit 1 on caregiver lexical alignment level (left) and syntactic alignment level (right). The plots are counter- factual in that they illustrate fictive data where we can vary only one of the factors at a time (while in real children they tend to co-vary): EL varies while VR and Socialization are kept at population average. The x- axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visit 2, 3, 4, 5, and 6. Note that the y-axes are on different scales to facilitate the evaluation of the effects.

Fig. 7.

Semantic alignment in real caregiver-child pairs (red for ASD, blue for TD) and 200 randomly sampled surrogate ones (green). Background transparencies display the average levels over time for each single real and surrogate caregiver. The x-axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visit 2, 3, 4, 5, and 6.

Fig. 8.

Panels illustrating the effects of the child’s EL, and socialization at visit 1 on caregiver semantic alignment. The plots are counterfactual in that they illustrate fictive children where we can vary only one of the factors at a time (while in real children they tend to co-vary): in the left column EL varies while VR and Socialization are kept at population average, in middle column VR varies, in the right column Socialization varies while EL and VR are kept at population average. The x-axis represents time from first visit, respectively 4, 8, 12 16 and 20 months later for visit 2, 3, 4, 5, and 6.