Front Matter

Ayah Zirikly, Andrew Yates, Bart Desmet, Molly Ireland, Steven Bedrick, Sean MacAvaney, Kfir Bar, Yaakov OphirOpenalex Unknown 2025 · DOI · OpenAlex

Authors: Ayah Zirikly, Andrew Yates, Bart Desmet, Molly Ireland, Steven Bedrick, Sean MacAvaney, Kfir Bar, Yaakov Ophir.
Source: Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2025).

2025.

Introduction

The 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2025), held in Albuquerque, New Mexico, represents a significant milestone in the intersection of Natural Language Processing (NLP) and mental health. As mental health conditions continue to be a leading cause of global disability—with an estimated economic impact of $16 trillion by 2030—the need for sophisticated, explainable, and clinically relevant computational tools has never been greater. This year's proceedings highlight a fundamental shift in the field: moving beyond simple binary classification toward a nuanced understanding of longitudinal mental health dynamics.

Beyond Static Classification

Historically, clinical NLP has focused on "static classification," attempting to label individuals with specific diagnoses (e.g., "depressed" vs. "not depressed") based on isolated snapshots of text. The organizers of CLPsych 2025 argue that this approach is insufficient for clinical practice. Mental health is inherently fluid; states fluctuate over time, and the underlying reasons for a prediction are often more valuable to a clinician than the prediction itself.

The workshop emphasizes the importance of capturing temporal dynamics and providing explainable evidence. Researchers are now prioritizing the extraction of specific behavioral markers and the summarization of a person's mental state over time, rather than relying on static diagnostic labels.

The CLPsych 2025 Shared Task: The MIND Approach

A central feature of the workshop was the 2025 Shared Task, which introduced a multi-task framework grounded in the transtheoretical MIND approach. This framework encourages models to analyze the interplay between adaptive (healthy) and maladaptive (unhealthy) self-states. The task was divided into four integrated components:

  1. Evidence Extraction: Identifying specific text spans that serve as indicators of adaptive or maladaptive states.
  2. Well-Being Score Prediction: Assigning a quantitative score (1–10) based on a user's psychological, social, and occupational functioning.
  3. Post-level Summarization: Generating natural language explanations of the mental states expressed within a single social media post.
  4. Timeline-level Summarization: Synthesizing how these states evolve across an entire longitudinal social media timeline.

Fourteen teams participated, utilizing a variety of methods ranging from traditional machine learning with domain-specific features to advanced Large Language Model (LLM) pipelines incorporating Retrieval-Augmented Generation (RAG) and few-shot prompting.

The shared task leveraged longitudinal social media timelines, continuing a trend toward modeling long-term behavioral patterns. Beyond the shared task, the proceedings include research utilizing diverse datasets, such as:

  • Semi-structured interviews for recognizing psychodynamic conflicts.
  • Artificial psychotherapy dialogues designed for empathy detection.
  • New benchmarks for "advanced mind-reading" (Theory of Mind) and the detection of thought disorders.

A key insight emerging from the workshop is that "bigger is not always better." While LLMs excel at high-level summarization, smaller neural language models often outperform their larger counterparts in specialized tasks like detecting thought disorders. Additionally, there is growing interest in "Synthetic Empathy," where LLMs generate psychotherapy dialogues to improve the training of counseling-support systems.

Results and Clinical Integration

The workshop maintained high academic standards, with a 48% acceptance rate (15 of 31 main papers). The success of the shared task demonstrated that LLMs can be effectively "prompt engineered" to provide nuanced clinical evidence extraction. Furthermore, the event successfully bridged the gap between disciplines, featuring a program committee and speakers—such as Zac Imel and Zohar Elyoseph—who represent both computational linguistics and clinical psychology.

Limitations and Future Directions

Despite these advancements, significant challenges remain:

  • Capturing Interplay: Models still struggle to fully capture the complex, dynamic interplay between different mental states.
  • Ethical Constraints: The sensitive nature of clinical data limits open research, necessitating a continued reliance on social media proxies.
  • Safety and Hallucination: In clinical settings, the risk of LLM hallucinations or biased explanations is a critical concern, requiring more robust evaluation frameworks.

Acknowledgments

The CLPsych 2025 workshop was supported by several organizations, including the Association for Computational Linguistics (ACL) and Ariel University.

Ariel University, a sponsor of the CLPsych 2025 workshop.

Summary updated

BibTeX
@inproceedings{zirikly2025front,
  author = {Ayah Zirikly and Andrew Yates and Bart Desmet and Molly Ireland and Steven Bedrick and Sean MacAvaney and Kfir Bar and Yaakov Ophir},
  title = {{Front Matter}},
  booktitle = {Openalex Unknown},
  year = {2025},
  doi = {10.18653/v1/2025.clpsych-1.0},
}
Figure