HARNN_ALemOA: A Novel Optimized Deep Learning Framework with Dynamic Prompting for Mental Health Classification Using Social Media Data

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Mental health problems adversely affect a person's emotional, social, psychological, and well-being. Individuals with mental health issues often do not seek professional help due to social stigma. Social media can serve as a vital tool for assessing a person’s mental health, and several methods have been proposed for classifying mental health from social media. Nevertheless, the prevailing methods mostly consider only the textual data without considering non-textual data, such as retweets and user engagement levels, which affect the reliability and accuracy of classification. Furthermore, the reviewed methods introduce challenges for text representation, multimodal information, and model optimization. Therefore, an effective model, namely Hierarchical Attention-based Recurrent Neural Network-based Aphid-Lemming Optimization Algorithm (HARNN_ALemOA) is proposed for classifying mental health. The objective is to improve classification by combining textual, contextual, and non-textual features and by optimizing the training of HARNN. Initially, input textual data is tokenized using Bidirectional Encoder Representations from Transformers (BERT), followed by feature extraction. Then, dynamic prompt engineering is employed to generate adaptive prompts tailored to the user's linguistic and emotional context, which are further refined using GPT to extract deeper contextual cues. Simultaneously, non-textual features are extracted. These features and contextual clues are combined and fed to HARNN with the Taylor-based Mean Bias Sigmoid Cross Entropy (TMSCE) loss for mental health classification. Further, HARNN is trained using ALemOA. Moreover, HARNN_ALemOA attained a better precision, F1-score, and recall of 97.979%, 97.633%, and 97.290%.