An Interlocutor-Modulated Attentional LSTM for Differentiating between Subgroups of Autism Spectrum Disorder

Yun-Shao Lin, Susan Shur-Fen Gau and Chi-Chun Lee

Abstract:

Recalling and discussing personal emotional experiences is one of the key procedures in assessing complex affect processing of individuals with Autism Spectrum Disorder (ASD). This procedure is a standard subpart of a diagnostic interview to assess ASD - the Autism Diagnostic Observation Schedule (ADOS). Previous work has demonstrated that the behavior features computed from this procedure in ADOS possess discriminative information between the three distinct ASD subgroups: Autistic Disorder (AD), High Functioning Autism (HFA) and Asperger Syndrome (AS). In this work, we propose an interlocutor-modulated attentional long short term memory network (IM-aLSTM) that models the ASD individual's acoustic features with a novel interlocutor-modulated attention mechanism. Our IM-aLSTM achieves ASD subgroup categorization accuracy of 66.5%, which is a 14% absolute improvement over baseline method on the same database. Our analyses further indicate that the attention weights are concentrated more on interaction segments where the ASD individual is being asked to recall and discuss his/her own negative emotional experiences.

---

Cite as: Lin, Y., Gau, S.S., Lee, C. (2018) An Interlocutor-Modulated Attentional LSTM for Differentiating between Subgroups of Autism Spectrum Disorder. Proc. Interspeech 2018, 2329-2333, DOI: 10.21437/Interspeech.2018-1288.

---

BiBTeX Entry:

@inproceedings{Lin2018,
author={Yun-Shao Lin and Susan Shur-Fen Gau and Chi-Chun Lee},
title={An Interlocutor-Modulated Attentional LSTM for Differentiating between Subgroups of Autism Spectrum Disorder},
year=2018,
booktitle={Proc. Interspeech 2018},
pages={2329--2333},
doi={10.21437/Interspeech.2018-1288},
url={http://dx.doi.org/10.21437/Interspeech.2018-1288} }