Abstract: Hiroki Tanaka, Sakriani Sakti, Graham Neubig, Tomoki Toda, Satoshi Nakamura. Proceedings of the Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality.
2014.
This research explores the potential of machine learning to bridge the gap between subjective clinical observation and objective data in diagnosing Autism Spectrum Disorder (ASD). By analyzing both the linguistic content ("what" is said) and acoustic properties ("how" it is said) of children's storytelling, the study demonstrates that automated systems can identify ASD markers with significant accuracy.
Addressing the Diagnostic Challenge
Diagnosing ASD traditionally relies on clinical assessments of social and communication deficits. While clinicians are trained to identify atypical prosody—the rhythm and melody of speech—or unusual word usage, these observations are often difficult to quantify and standardize. Previous research has frequently examined speech or language in isolation. This paper proposes an integrated approach, combining acoustic signals (such as pitch and voice quality) with linguistic patterns (word categories) to create a robust, automated screening tool that remains effective across different languages and cultures.
Methodology and Data Collection
The researchers developed a multi-layered pipeline for feature extraction and classification:
- Feature Extraction:
- Linguistic: The Linguistic Inquiry and Word Count (LIWC) tool was used to categorize words into groups such as social processes, affective words, and cognitive states.
- Acoustic: The Snack sound toolkit measured fundamental frequency (pitch), intensity (volume), and voice quality (formants and harmonics).
- Narrative-Level: Turn-taking dynamics were quantified, specifically measuring the duration of pauses before a child began speaking after a parent's prompt.
- Statistical Analysis: T-tests, Principal Component Analysis (PCA), and Factor Analysis were employed to identify features contributing most to the variance between ASD and typically developing (TD) groups.
- Machine Learning: Naive Bayes (NB) and Support Vector Machines (SVM) were tested for utterance-level classification, while a K-Nearest Neighbor (K-NN) approach was used for full narrative-level analysis.
The study utilized a newly collected dataset of interactions between Japanese children (4 with ASD, 2 TD) and their parents across various tasks, including free play and storytelling. To ensure cross-cultural validity, these findings were compared against the USC Rachel corpus, which contains similar interactions with American children.
Key Insights: Vocabulary and Prosody
The analysis revealed a distinct "social vocabulary gap." Typically developing children used significantly more words related to social processes, emotions, and cognitive states (e.g., "know," "think"). In contrast, children with ASD relied more heavily on fillers and simple assent words.
Furthermore, the study quantified the clinical observation of "monotonous" speech in ASD. This was reflected in significantly lower standard deviation and coefficient of variation in fundamental frequency (pitch) compared to the TD group.
The Power of the Pause
One of the most striking findings was the difference in response latency. Children with ASD took significantly longer to respond to their parents, particularly following open-ended questions. This suggests that the cognitive load of processing social cues and formulating a narrative response is a measurable diagnostic marker. This "long pause" marker was consistent across both Japanese and American cohorts, suggesting a potentially universal social-communication marker for ASD.
Classification Results
The integrated model showed high efficacy in distinguishing between the two groups:
- Utterance-Level: Using an SVM with a selected set of statistically significant features, the system achieved 71.9% accuracy.
- Narrative-Level: When aggregating features across an entire 5-minute story, the K-NN classifier achieved 100% accuracy on the study's sample.
- Feature Dominance: Decision tree analysis indicated that acoustic features, particularly pitch variation (represented by 'fsd' and 'fcov'), were often the most effective primary discriminators.
Limitations and Future Directions
While the results are promising, the authors acknowledge the primary limitation of a small sample size (six children in the Japanese cohort). The 100% narrative-level accuracy serves as a proof of concept rather than a definitive clinical result. Future work will focus on scaling the study to larger, more diverse datasets to validate these features and developing real-time dialogue systems for automated screening or social skills coaching.
Summary updated