BERT- BiLSTM-Caps Language Model for Screening of Children's Severe Mental Retardation

Abstract

Current screening for mental retardation in children typically uses a clinical intelligence assessment, an exercise that often relies on the diagnosis of professional psychiatrists. Studies have shown that mental retardation is usually accompanied by symptoms such as language impairment and communication difficulties. In this work, we propose a language-based model (BERT-BiLSTM-Caps) to assist in screening for mental retardation in children. Specifically, we collect audio of multiple child participants' intelligence assessment interview sessions and convert it into transcripts for training the neural network. We use a pre-trained language model BERT to obtain an encoded representation of the interview text, and BiLSTM is used to learn global features. Innovatively, we introduce a capsule network with a dynamic routing mechanism to further extract critical features. Our proposed model obtains an accuracy of 89.1 % on the children interview dataset. The experimental results show that it is potential for screening children's mental retardation by analyzing their language patterns.