Corpus-Adaptive Symptom Prototype Learning for Social-Media Depression Detection

Abstract

Psychiatric-scale-guided screening has shown promise for social-media depression detection, but existing methods largely treat clinical symptom templates as fixed inputs and focus on improving screening or classification architectures. We instead investigate whether the symptom representation itself can be adapted to social-media language while retaining psychiatric priors. Fixed clinical templates exhibit three key mismatches with online expression: formal versus colloquial language, limited coverage of platform-specific signals, and linguistic distance that can weaken semantic matching. To address these mismatches, we propose Corpus-Adaptive Symptom Prototype Learning, which derives complementary symptom prototypes from depressed users’ training posts through semantic clustering and symptom extraction, refinement and deduplication, and social-media-style rewriting. The learned prototypes are combined with existing psychiatric templates and integrated in a plug-and-play manner without modifying downstream detection architectures. Experiments on eRisk2017 and eRisk2018 across five structurally diverse detection backbones show that the proposed representation consistently improves F1 and recall across all ten dataset–backbone combinations. Ablation results further show that the gains arise from corpus-derived symptom information rather than LLM-generated reformulations alone, and all three prototype-construction stages contribute to performance. These results demonstrate that adapting symptom representations, rather than only detection architectures, provides an effective and architecture-agnostic extension to psychiatric-scale-guided screening.