Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology

Elizabeth Shriberg, Ayah Zirikly, Dana Atzil–Slonim, Maria Liakata, Steven Bedrick, Bart Desmet, Molly Ireland, Andrew Lee, Sean MacAvaney, Matthew Purver, Rebecca Resnik, Andrew Yates, Mashrura Tasnim, Malikeh Ehghaghi, Brian Diep, Jekaterina Novikova, Harold Goodglass, Edith Kaplan, Sandra Weıntraub, Anupama Ray, Siddharth Krishna Kumar, Rutvik Reddy, Pre- Rana Mukherjee, Ritu Garg, Multi, David Bloom, Elizabeth Cafiero, Eva Jan-Llopis, Shafika Abrahams-Gessel, Lakshmi Bloom, Sana Fathima, Andrea Feigl, Tom Gaziano, Ali Hamandi, Mona Mowafi, Daniel C Ehlman, Sara David C Sheridan, Rebecca Grusing, Amber Marshall, Adrienne Lin, Robert Hughes, Zane Hendrickson, Horowitz, Adam Tsakalidis, Jenny Chim, Iman Bilal, Dana Atzil-Slonim, Federico Nanni, Philip Resnik, Manas Gaur, Kaushik Roy, Becky Inkster, Jeff Leintz, Noori Akhtar-Danesh, Janet Landeen, Murray Alpert, Enrique Pouget, Raul Silva, Rg Bachu, B Kopparthi, Adapa, Barkana, Megan Barker, Breanne Young, Gail Robinson, Aaron T Beck, Robert Steer, Margery Carbin, Mehnaz Adnan, Jim Warren, Martin Orr, Awais Aftab, Elizabeth Ryznar, Chittaranjan Andrade, Angele Marie, Beausoleil, Adrian Benton, Glen Coppersmith, Mark Dredze, German Berrios, Ananya Bhatia-Lin, Alexandra Boon-Dooley, Michelle Roberts, Caroline Pronai, Dylan Fisher, Lea Parker, Allison Engstrom, Leah Ingraham, Doyanne Darnell, M David, Blei, Michael Ng, Jordan, Virginia Braun, Victoria ClarkeOpenalex Unknown 2022 · DOI · OpenAlex

Abstract

Mental health is a pressing global concern, with conditions among the leading causes of disability. In the U.S., suicide ranks as a top cause of death across several age groups, a crisis exacerbated by the COVID-19 pandemic. The Eighth Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2022) adopted the theme "mental health in the face of change," focusing on natural language processing (NLP) solutions that are explainable, fair, and capable of capturing mood changes over longitudinal data.

CLPsych 2022: Mental Health in the Face of Change

The 8th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2022) marked a significant shift in the field, moving from static "black-box" classification toward longitudinal, explainable modeling of mental health. By introducing new datasets for speech and social media timelines, the workshop highlighted how NLP can track "Moments of Change" in mood and suicidality, effectively bridging the gap between raw data and clinical formulation.

Addressing the Dynamic Nature of Mental Health

Mental health is inherently dynamic, yet historical NLP research has often treated it as a static problem—labeling a user as "depressed" based on a single snapshot. This fails to account for symptom fluctuation or the impact of treatment over time. Furthermore, clinical adoption is hindered by a lack of explainability. CLPsych 2022 addressed these gaps by emphasizing longitudinal tracking and clinical grounding using established tools like the PHQ-9.

Technical Methods and Innovations

The workshop showcased diverse technical approaches to capture these temporal changes:

  • Longitudinal Modeling: Participants in the Shared Task identified sudden mood shifts ("Switches") and gradual declines ("Escalations") in Reddit timelines using BiLSTMs and emotionally-informed seq2seq architectures.
  • Clinically Grounded Auxiliary Tasks: Researchers utilized PHQ-9 categories as auxiliary labels in multi-task learning, forcing models to learn features aligned with clinical symptoms and improving interpretability.
  • Contextual Graph Neural Networks (GNNs): Suicidality detection was enhanced by modeling relationships between posts and suicide dictionaries as heterogeneous graphs to better distinguish risk context.
  • Morphosyntactic Masking: A novel approach involved masking specific parts of speech to determine which linguistic categories, such as nouns in schizophrenia diagnosis, are most critical for clinical signal.

New Benchmarks and Corpora

Several new benchmarks were introduced to support this research:

  • DEPAC: An audio dataset for depression and anxiety detection featuring acoustic and linguistic features from 571 subjects.
  • PRIMATE: Designed to train models to recognize which PHQ-9 questions are "answered" by a user's initial post.
  • Shared Task Dataset: Curated longitudinal Reddit posts re-annotated for "Moments of Change."
  • Talkspace Corpus: 24 million words from real-world text-based therapy sessions used to extract emotional signals.

Key Insights and Results

Research presented at the workshop revealed that linguistic evidence of depression in self-disclosed users decreases significantly over five years, suggesting that static labels in old datasets may introduce noise. Additionally, transfer learning showed that models trained on one condition (e.g., stress) could predict others (e.g., suicidal ideation) with F1 scores up to 0.80, indicating a shared linguistic footprint across mental health conditions.

UMIACS Logo

Limitations and Future Directions

Challenges remain regarding data scarcity, privacy, and demographic bias. Generative models for follow-up questions still risk "hallucination," requiring human-in-the-loop oversight. The field is now moving toward "digital guardians"—tools providing real-time support while adhering to strict medicolegal and ethical standards.

American Association of Suicidology Logo

Summary updated

Abstract

Mental health is a pressing concern. Worldwide, mental health conditions are among the leading causes of disability In the U.S. in 2020, suicide was in the top nine leading causes of death for people ages 10-64, and the second leading cause of death for people ages 10-14 and 25-34 [2]. Over the past several years, COVID-19 has created additional challenges to mental health. For instance, Sheridan et al. Furthermore, U.S. Surgeon General Vivek Murthy in 2021 called for a nationwide response to the mental health crisis that youth especially are facing during the pandemic For the Eighth Workshop on Computational Linguistics and Clinical Psychology (CLPsych), we adopt the theme "mental health in the face of change". This includes the kind of aspects natural language processing technologies need to address to deliver explainable and fair solutions that can be integrated in the clinical setting. Additionally, how these solutions can capture changes in mood over longitudinal and temporal data, which has been the focus of this year's shared task. CLPsych was a hybrid workshop that accommodated both in-person and remote participation. It was collocated with NAACL'22, which took place in Seattle, Washington, USA on July 15 th , 2022. Since 2014, CLPsych has been successful in bringing together people from different backgrounds (e.g. mental health experts, clinicians, and computational linguists), to share and discuss their work and results. Its central goal is to build bridges so that these different disciplines can integrate to improve our understanding of mental health issues, and to deliver better mental health treatments and diagnoses to everybody. CLPsych'22 included a shared task that focused on using longitudinal data to understand mood changes and relate them to risk assessment for suicidality. The shared task was organized by Adam Tsakalidis, Federico Nanni, and Maria Liakata. The overview of the shared task in this volume [6] discusses the tasks, team approaches and results, and lessons learned. Our program committee included mental health and technological experts, in order to provide all the papers with more informative feedback that address both aspects. CLPsych'22 received a total of 23 papers for the main workshop, of which 15 were accepted; all 9 submitted shared task papers were also accepted. The organizing committee, with the help of the program committee scores, and feedback chose seven main workshop papers and two shared task papers as oral presentations, and the rest were presented in the poster session. CLPsych'22 also hosted excellent invited speakers and panelists.
BibTeX
@inproceedings{shriberg2022proceedings,
  author = {Elizabeth Shriberg and Ayah Zirikly and Dana Atzil–Slonim and Maria Liakata and Steven Bedrick and Bart Desmet and Molly Ireland and Andrew Lee and Sean MacAvaney and Matthew Purver and Rebecca Resnik and Andrew Yates and Mashrura Tasnim and Malikeh Ehghaghi and Brian Diep and Jekaterina Novikova and Harold Goodglass and Edith Kaplan and Sandra Weıntraub and Anupama Ray and Siddharth Krishna Kumar and Rutvik Reddy and Pre- Rana Mukherjee and Ritu Garg and Ritu Garg and Multi and David Bloom and Elizabeth Cafiero and Eva Jan-Llopis and Shafika Abrahams-Gessel and Lakshmi Bloom and Sana Fathima and Andrea Feigl and Tom Gaziano and Ali Hamandi and Mona Mowafi and Daniel C Ehlman and Sara David C Sheridan and Rebecca Grusing and Amber Marshall and Adrienne Lin and Robert Hughes and Zane Hendrickson and Horowitz and Adam Tsakalidis and Jenny Chim and Iman Bilal and Ayah Zirikly and Dana Atzil-Slonim and Federico Nanni and Philip Resnik and Manas Gaur and Kaushik Roy and Becky Inkster and Jeff Leintz and Maria Liakata and Noori Akhtar-Danesh and Janet Landeen and Murray Alpert and Enrique Pouget and Raul Silva and Rg Bachu and B Kopparthi and Adapa and Barkana and Megan Barker and Breanne Young and Gail Robinson and Aaron T Beck and Robert Steer and Margery Carbin and Mehnaz Adnan and Jim Warren and Martin Orr and Awais Aftab and Elizabeth Ryznar and Chittaranjan Andrade and Angele Marie and Beausoleil and Adrian Benton and Glen Coppersmith and Mark Dredze and German Berrios and Ananya Bhatia-Lin and Alexandra Boon-Dooley and Michelle Roberts and Caroline Pronai and Dylan Fisher and Lea Parker and Allison Engstrom and Leah Ingraham and Doyanne Darnell and M David and Blei and Andrew David M Blei and Michael Ng and Jordan and Virginia Braun and Victoria Clarke and Virginia Braun},
  title = {{Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology}},
  booktitle = {Openalex Unknown},
  year = {2022},
  doi = {10.18653/v1/2022.clpsych-1},
}
Figure