Year: 2026
Type: conference-paper
Abstract: Targeting the CLPsych 2026 (Ali et al., 2026) shared task, this study proposes a unified framework based on three complementary strategies.We implement a pipeline integrating post-level psychological state recognition, timeline-level change detection, and evidence-based clinical summarization.For Task 1, we evaluate three paradigms for extracting ABCD elements: zero-shot prompting, Qwen-3 fine-tuning, and TextCNN-based hierarchical classification.For Task 2, we design a cascaded temporal model that combines Mental-BERT features, log-time encoding, and Bi-LSTM to capture mental state switches and escalations under irregular posting intervals.For Task 3, we adopt a dynamic retrieval-augmented generation strategy to produce structured clinical summaries.Experimental results show that our system ranks second on the official leaderboard for Task 2, demonstrating the effectiveness of time-aware modeling for change detection.Error analysis further reveals over-prediction bias and valence inversion in Tasks 1 and 3, highlighting pipeline error propagation as a key challenge for future work.
Authors: Mengjia Zhang, Rui Chen, Haonan Xiao, Yi Yang
Institution: Chongqing Technology and Business University
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166826388