Linguistic Features of Clients and Counselors for Early Detection of Mental Health Issues in Online Text-based Counseling.

Online counseling is essential for overcoming mobility restrictions, schedule limitations, and mental health stigma. However, government counseling offices are being inun-dated with consultations for which non-mental health supports are targeted. Therefore, we aim to create a classification model that classifies whether the clients have mental health issues or other issues. We expect to support counselors by presenting the classification results. We conducted the first automatic detection of clients who might be suffering from mental health issues and used almost 1000 actual counseling sessions for our machine learning framework. We achieved an F1-score of 0.646 by classifying dialogue sessions using features such as frequency-inverse, document frequency, document embedding of a large-scale language model, linguistic inquiry and word count, topic modeling, and statistics of dialogue sentences. In addition, we performed dimensionality reduction with principal component analysis. We also conducted evaluation experiments using dialogue sentences from the beginning to the middle of sessions as input and clarified the relationship between the number of messages in the dialogues and the transition in the classification performance. We also identified the words that contribute to detecting mental health issues for each client and counselor. Clinical relevance-This study makes it possible to detect the trends identified in a client's anxieties during counseling. Our findings are critical for designing systems that assist counselors.