TL;DR
This document presents the proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2025), focusing on the theme "understanding the mental health state – going beyond classification." It highlights advancements in using LLMs for mental health analysis, including a shared task that explored capturing mental health dynamics from social media timelines through a novel multi-task framework.
What Problem Does This Paper Address?
Mental health conditions are a significant global burden, contributing to disability and economic strain. The COVID-19 pandemic has further exacerbated these challenges, with concerning increases in mental health crises, particularly among younger populations. While traditional approaches to mental health analysis often rely on classification, there's a growing need to move beyond simple categorization towards a deeper understanding of mental health states. This is where computational linguistics, particularly large language models (LLMs), shows immense promise. LLMs can not only aid in diagnosis but also provide nuanced explanations for prediction outcomes, offering a richer insight into an individual's mental state. CLPsych aims to foster interdisciplinary collaboration between mental health experts and computational linguists to address these critical issues and improve mental health treatments and diagnoses.
Methods
The CLPsych 2025 workshop featured a shared task that introduced a novel multi-task framework grounded in the transtheoretical MIND approach. This framework builds upon previous work in longitudinal modeling and focuses on capturing mental health dynamics from social media timelines. The shared task involved four subtasks:
- Evidence Extraction: Identifying text spans that reflect adaptive or maladaptive self-states.
- Well-Being Score Prediction: Assigning a score from 1 to 10 based on social, occupational, and psychological functioning.
- Post-level Summarization: Summarizing the interplay of adaptive and maladaptive states within individual posts.
- Timeline-level Summarization: Capturing the temporal dynamics of self-states across a series of posts in a timeline.
The approaches submitted by the 14 participating teams ranged from traditional machine learning with domain-specific features to sophisticated LLM pipelines incorporating demonstration and retrieval.
Datasets
Not explicitly detailed in the provided pages, but the shared task focused on analyzing social media timelines. The papers presented likely utilized various social media datasets, potentially including publicly available datasets or proprietary data collected for research purposes. The CLPsych 2025 shared task itself serves as a benchmark for evaluating methods in capturing mental health dynamics from social media.
Key Insights and Ideas
- Beyond Classification: The overarching theme emphasizes the shift from simply classifying mental health states to understanding their nuances and dynamics. This is crucial for developing more effective interventions.
- LLMs as Powerful Tools: LLMs are demonstrating significant potential in mental health, not just for diagnosis but for providing explanations and capturing complex patterns in text.
- Longitudinal Analysis is Key: The focus on social media timelines highlights the importance of understanding how mental health states evolve over time, rather than just analyzing isolated posts.
- Interdisciplinary Collaboration is Essential: The workshop's success relies on bringing together experts from computational linguistics and clinical psychology to ensure that computational methods are both technically sound and clinically relevant.
- Hybrid Approaches are Promising: The shared task saw a range of approaches, from traditional ML to LLM-based solutions, suggesting that hybrid methods might offer the best of both worlds.
Key Results
- 14 teams completed the CLPsych 2025 shared task, showcasing diverse methodologies.
- The results from the shared task indicate the complexity of capturing mental health states beyond static classification.
- The workshop received 31 papers for the main track, with 15 accepted, and all 11 submitted shared task papers were accepted.
- 6 main workshop papers and 4 shared task papers were selected for oral presentations, with the rest presented as posters.
Limitations and Future Work
The proceedings highlight the complexity of capturing mental health states, suggesting that future work should continue to explore more sophisticated methods for dynamic analysis. The paper implicitly suggests that further research is needed to develop validated, dimensional, and transdiagnostic approaches to mental health assessment using computational methods. The success of the shared task also points towards future research in leveraging LLMs for more nuanced understanding and summarization of clinical texts and social media data related to mental health.
Summary updated