Aphasia Detection for Cantonese-Speaking and Mandarin-Speaking Patients Using Pre-Trained Language Models

Automatic analysis of aphasic speech based on speech technology has been extensively investigated in recent years, but there has been a few studies on Chinese languages. In this paper, we focus on automatic aphasia detection for Cantonese-and Mandarin-speaking patients using state-of-the-art pre-trained language models that support both traditional and simplified Chinese. Given speech transcriptions of subjects, pre-trained language models are used in two ways: 1) pre-trained language model derived embeddings followed by a classifier; 2) pre-trained language model fine-tuned for aphasia detection task. Both approaches are demonstrated to outperform baseline models using acoustic features and static word embeddings. The best accuracy is obtained with fine-tuned BERT models, achieving 0.98 and 0.94 for Cantonese-speaking and Mandarin-speaking subjects respectively. We also investigate the feasibility of applying the cross-lingual pre-trained language model fine-tuned by aphasia detection task for Cantonese-speaking subjects to Mandarin-speaking subjects with limited data. The promising results will hopefully make it possible to perform detection on those low-resource pathological speech which is difficult to implement a specific detection system.