Automatic Detection of Public Speaking Anxiety Through Linguistic Analysis

The automatic detection of Public Speaking Anxiety (PSA) can facilitate prompt assistance for persons facing anxiety, enhancing communicative abilities and mental health. While prior research investigated acoustics and visual features to automatically detect PSA, this study investigates whether the textual modality can benefit PSA detection as well. We used Bidirectional Encoder Representations from Transformers (BERT) and Linguistic Inquiry and Word Count (LIWC) text features with several machine learning models, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). The SVM model using BERT features on text chunks produced the best results, with an F1 score of 0.76. Although the experiments employed a small dataset, the findings indicate that text features may play an important role in identifying PSA. In future work, we aim to investigate the respective information contained in textual vs acoustic features using multimodal approaches such as LSTM models. Our code available at:https://github.com/M23A?tab=repositories