Deep Author Profiling (Deep-AP) is a unified multi-task deep learning architecture that jointly infers age and gender at the author level and depression at the publication level from social media text. Existing approaches commonly treat Author Profiling (AP) and Depression Detection (DP) as independent problems, which increases model redundancy and computational cost and limits scalable deployment. To address this gap, Deep-AP adopts a shared-encoder, multi-head formulation that enables parameter sharing across heterogeneous tasks while preserving task-specific outputs. The study follows a systematic methodology that combines large-scale Reddit data collection, automatic silver-standard annotation via regular expressions and subreddit supervision, and user-level aggregation of multiple posts to support author-centric inference. Deep-AP processes up to twelve publications per user through a single backbone encoder (GRU, BiGRU, BERT variants, or LLaMA-3.2), producing post-level representations for DP and aggregated user-level representations for demographic prediction. We also introduce Deep Author Profiling Username Enhanced (Deep-AP-uname), which incorporates username embeddings as auxiliary input to enrich demographic inference. Experiments across backbone models show that BERT-large-based Deep-AP achieves F1-scores of 0.914, 0.717, 0.881 for gender, age, and depression respectively . These findings demonstrate that joint learning can deliver strong predictive performance while improving practicality, reducing inference time by 66% and memory usage by 40% compared to parallel single-task BERT-large models.
Oulahbib et al. (2026) studied this question.