DiscRST-HTT: Hierarchical Multimodal Transformer with Adaptive Gating for Discourse-Aware Depression Detection

Abstract

Recently, depression detection through social media activity has emerged as an effective and gain insights to understand the mental health condition of a user. However, traditional transformer-based models primarily focus on textual semantics and neglect the temporal evolution of behavior, which limits their adaptability and reliability. To address these challenges, a Discourse-aware Rhetorical Structure Theory with Hierarchical Temporal Transformer (DiscRST-HTT) is proposed for integrating contextual, rhetorical, and behavioral cues for comprehensive depression analysis. The contextual semantics are captured through a Mental Bidirectional Encoder Representations from Transformer (MentalBERT), while rhetorical embeddings are derived from a Neural Rhetorical Structure Theory (NeuralRST) model. In contrast, the model learns behavioral embeddings, which are captured based on posting frequency. These embeddings are fused through a multimodal adaptive gating mechanism (MAGM) by dynamically balancing modality contributions. Then, the fused embeddings are processed by a hierarchical temporal transformer (HTT) model, which captures both intra-post and inter-post dependencies with an ordinal regression head to learn ordered depression severity representations. While the final evaluation is conducted through binary depression classification derived from threshold ordinal output. From the results, the proposed DiscRST-HTT attains better binary classification performance derived from ordinal severity predictions achieving accuracy of 98.75%, precision 98.34%, recall 99.05%, and F1-score 98.78% on the Reddit dataset when compared to the existing BERT with convolutional neural network (BERT-CNN) model.