Recent years have witnessed a escalating prevalence and severity of mental health challenges among the general populace, a trend corroborated by a multitude of research studies. Traditional diagnostic methodologies for mental illness typically involve hospital-based psychological counseling. However, certain individuals confront deeply ingrained stigmatized views, which can impede the timely detection and intervention of mental health issues due to their nuanced and often latent nature. In response to these challenges, this investigation introduces an innovative text classification model termed RoBERTa-wwm-CBA (RoBERTa-wwm-CNN-BiLSTM-Attention), which combines the powerful feature extraction capabilities of the RoBERTa-wwm language model with a hybrid neural network architecture. This model is designed to effectively identify potential psychological distress in individuals seeking counseling by mining text features across various levels of granularity. The RoBERTa-wwm-CBA model leverages the RoBERTa-wwm’s pre-trained expertise to seize the nuanced features within text data. It incorporates a CNN layer to extract local semantic features from the text, integrates these with broader contextual information through a BiLSTM layer, and applies an attention mechanism to ascribe differential importance to different features. This facilitates the derivation of a holistic text vector representation. Empirical evaluations on the Emotional First Aid Dataset have revealed that the RoBERTa-wwm-CBA model achieves superior performance in terms of accuracy, precision, recall, and F1 score, substantially enhancing the efficiency and accuracy of mental illness recognition.