Result: Computationally characterizing communicative content and context in autistic children

Pragmatic language difficulties are common among autistic children, and assessment of pragmatic language skills over time is an important predictor of quality of life outcomes during adulthood. Current metrics for pragmatic language are qualitative in design and are expensive in terms of time and resources. With the use of Natural Language Processing (NLP) methods, robust measures of pragmatic language features can be obtained in an automated, reliable, and relatively inexpensive fashion. Such metrics can be used to augment traditional pragmatic language assessments. Improving our understanding of how autistic individuals use language not only helps us learn how to become better conversational partners ourselves, but also enables us to build language tools that accommodate for pragmatic language differences. In this dissertation, we leverage traditional statistical methods to adapt and augment established NLP techniques to investigate three areas of pragmatic language that autistic children are known to have difficulty with. We use a corpus of transcribed Autism Diagnostic Observation Schedule (ADOS) sessions for 117 autistic children (98 male) and 65 Typically Developing (TD) children (37 male), aged 4 to 15 years old. We first compare how autistic children use the fillers "um" and "uh" differently than their TD peers during conversations. After controlling for age, sex, and IQ, we found that autistic children used less "um" frequently than their TD peers and that structural language scores predicted "um" usage while social affect and pragmatic language scores did not. Next, we investigate differences in topic maintenance ability. We present a novel statistical approach for investigating group difference in the document-topic distribution vectors created by Latent Dirichlet Allocation (LDA). After transforming the vectors using Aitchison geometry, we use multivariate analysis of variance (MANOVA) to compare sample means and calculate effect size using partial eta-squared. We validate our method on a subset of the 20Newsgroup corpus and then apply our method to our clinical corpus. We found that the topic distribution vectors of autistic children significantly differed from those of TD children when responding to questions about social difficulties. Lastly, we investigate differences in backchannel usage (i.e., "right", "okay", "uhhuh") between autistic and TD children. After adjusting for age, sex, and IQ, we found that autistic children used less backchannels than their TD peers and were less likely to produce a backchannel with a greater overlap length.