Accurate Detection of Dementia from Speech Transcripts Using RoBERTa Model
Lovro Matosevic, Alan Jović · 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO) · 2022
Dementia is a serious disease that is very common in the elderly population. Automatic detection of dementia is a difficult task that may involve the analysis of acoustic features of speech, linguistic features of transcripts, and mental state exams. In this work, we explore the limits of using speech transcripts from doctor-patient conversations to detect dementia. The dataset is prepared from Pitt corpus, which is a part of DementiaBank, a shared database of multimedia interactions for studying communication in dementia. We use a sophisticated natural language processing approach, namely RoBERTa, which addresses the problem by using transformers and self-attention mechanism. We compare RoBERTa with a baseline BERT model. We show that dementia detection using well-prepared speech transcripts alone can lead to detection rates above 90% for RoBERTa model in a near-balanced dataset, outperforming the baseline model.