Hierarchical Multi-Stage Modeling of Adaptive and Maladaptive Self-States in Social Media Timelines
Year: 2026
Type: conference-paper
Abstract: We address the CLPsych 2026 Shared Task on modeling psychological self-states from longitudinal social media data.We propose (i) a hierarchical multi-stage framework that integrates a multi-task transformer encoder and (ii) a four stage instruction-tuned large language model finetuning pipeline for subelement classification, presence estimation, and evidence extraction.Our approach incorporates elementconditioned label masking and cross-stage encoder transfer, enabling structured prediction aligned with the ABCD psychological framework.Experiments show improvements over the baseline on the development setup, with RoBERTa achieving an 8.3% gain in macro-F1 and improved root mean square error (RMSE), while a fine-tuned Qwen3 model attains the best overall performance.These results demonstrate the effectiveness of combining hierarchical multi-task learning with structured generation for interpretable mental health analysis.All code for this task is publicly available in a GitHub repository 1 .
Authors: Abir Naskar, Mike Conway
Institution: The University of Melbourne
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166858711