[3A1-GS-10] AI application:

[3A1-GS-10-02] Automatic Psychological Counseling Classification Using BERT-Based Models and LLM-Driven Data Augmentation

〇Chi Zhang¹, Ryo Sekizaki², Shengzhou Yi¹, Toshihiko Yamasaki¹ (1. The University of Tokyo, 2. Welcome to talk Co. Ltd.)

Keywords:Mental Health, Large Language Model, Natural Language Processing

Psychological counseling has become an essential service in real life. With the volume of counseling records growing rapidly, it becomes infeasible for experts to manually assess every submission. This escalating demand highlights the pressing need to develop language models capable of automatically classifying counseling records based on their severity. In this study, we develop a Bidirectional Encoder Representation from Transformers (BERT)-based model to enhance the effectiveness of counseling severity classification. The model's performance is limited by the small dataset size, particularly in the critical high-severity category. To address this, we utilize GPT-4o to generate synthetic counseling records for these underrepresented categories, effectively augmenting the dataset. Furthermore, we introduce an undersampling method to rebalance the distribution of high- and low-severity categories. The proposed approaches significantly improve the model's performance, achieving an accuracy of 80.8% in classifying high-severity counseling records.