Center for Spoken Language Understanding
Oregon Health & Science University
Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by impaired social communication, and restricted, repetitive patterns of behavior and interest. These two core symptoms can appear at the language level and result in problems such as using inappropriate words, idiosyncratic language, topic repetition, and lack of conversational responsiveness. Existing methods for the evaluation of language in ASD are mainly based on subjective parental and clinical reports. In this thesis, we propose fast, objective, scalable, automatic analysis of these interrelated aspects of ASD language, utilizing computational methods for natural language processing based on unannotated verbatim transcripts of conversations. We first apply word ranking and distributional semantic models to automatically determine off-topic lexical content in children’s narrative retellings. Our classification of unexpected words is sufficiently accurate to distinguish the retellings of children with autism from those with typical development (TD). Second, we utilize semantic similarity measures to identify idiosyncratic topic digressions expressed in narratives. Our findings indicate that TD children tend to use similar words and semantic concepts when retelling the same narrative, while children with ASD use different words and concepts that are potentially related to their individual topics of interest. Third, we try to quantify restrictive and repetitive interests and topic repetition in spontaneous conversations of autistic children. Using various similarity measures, we show that the children with ASD have significantly higher ratio of semantically overlapping dialogue turns compared to their TD peers, as a result of higher topic perseveration in their conversations. Finally, we focus on social communication and interaction in children with ASD and we analyze their question responsiveness as well as the use of discourse markers and acknowledgments in various contexts. Our findings suggest that the ASD children are less responsive compared to TD children, and have problems in the appropriate use of conversational cues. Our proposed methods and results in this thesis underscore the potential of automated natural language processing techniques for improving the understanding of the prevalence and diagnostic significance of language use in ASD.
School of Medicine
Rouhizadeh, Masoud, "Computational analysis of language use in autism" (2015). Scholar Archive. 3732.