GEM: graph attention encoder for multi-task depression severity detection in multi-party conversations

Depression detection (DD) using online text, including chat conversations, is an emerging research area. Multi-party conversation (MPC) analysis focuses on uncovering complex discourse-level relationships between utterances, facilitating tasks like emotion recognition (ER) of interlocutors. While graph attention networks (GATs) have demonstrated state-of-the-art (SOTA) performance in ER tasks for MPCs, their application to depression severity classification (DSC) has not been explored. To address this gap, we propose GEM, a GAT-based encoder designed for DSC in MPCs, which incorporates root- and sub-level utterances to introduce hierarchy and depth into conversation representations. GEM uses multiple edge types to preserve relationships between utterances, speaker-specific details, and temporal information. This enables the incorporation of discourse relations across utterances, the interpretation of individual behaviors, and effective tracking of mood shifts. A multi-task learning pipeline is utilized to jointly model both DD and DSC. Extensive experiments on tasks including depressed utterance detection (DUD), depressed interlocutor recognition (DIR), and DSC demonstrate that GEM achieves SOTA performance. Notable improvements include a 10.22% increase in recall for DUD, along with 10.92 and 5.07% improvements in F1 scores for DIR and DSC, respectively, across the eRisk 18 T2, Twitter Depression 2022, and DEPTWEET corpora. These results highlight GEM’s potential as a screening tool for DD.