The Language of Trauma: Modeling Traumatic Event Descriptions Across Domains with Explainable AI

Miriam Schirmer, Tobias Leemann, Gjergji Kasneci, Jürgen Pfeffer, David Jurgens

Introduction

Psychological trauma is a deeply subjective experience, and its linguistic expression varies significantly depending on the social and situational context. A witness testifying in a genocide tribunal employs a formal, descriptive register that differs vastly from the peer-to-peer support language found on Reddit or the professional dialogue of a counseling session. Most existing Natural Language Processing (NLP) models for mental health are confined to single domains, often failing when applied to new contexts. Furthermore, these "black box" models frequently lack the interpretability required for clinical or sensitive applications.

To address these challenges, researchers have introduced TRACE (Trauma Event Recognition Across Contextual Environments), a cross-domain framework designed to identify traumatic event descriptions while providing explainable insights into the underlying linguistic markers.

Methodology and Data Sources

The TRACE framework utilizes a multi-tiered approach, combining fine-tuned transformer models with an Explainable AI (XAI) stack to ensure both high performance and transparency.

The TRACE framework overview, illustrating the pipeline from diverse data sources to binary classification and XAI-driven characteristic identification.

The TRACE Dataset

The study aggregates four distinct data sources to test the model's robustness:

  1. Genocide Transcript Corpus (GTC): Formal witness statements from international criminal tribunals (e.g., Rwanda, Yugoslavia).
  2. PTSD Subreddit: Informal, personal narratives and peer support from online communities.
  3. Mental Health Counseling Conversations: Professional Q&A sessions between users and psychologists.
  4. Incel Posts (Control Group): A dataset characterized by aggressive and violent rhetoric, used to test if the model can distinguish between general hostility and actual traumatic event descriptions.

Modeling and Explainability

The researchers compared traditional baselines (Naive Bayes, Logistic Regression) against deep learning architectures, including fine-tuned BERT, RoBERTa, and Large Language Models like GPT-4. To interpret the models' decisions, they employed SHAP for local feature attribution and SLALOM to distinguish between a token's value (class contribution) and its importance (model reliance).

Key Results and Cross-Domain Transferability

A primary goal of the study was to determine if a "universal" language of trauma exists across these disparate domains. The results indicate a surprising level of transferability, particularly for models trained on the extreme narratives of the Genocide Transcript Corpus.

Heatmap showing cross-domain performance (F1-scores) across different training and testing datasets, highlighting the transferability of trauma detection models.

Fine-tuned RoBERTa emerged as the top performer, achieving F1-scores of 0.74 on the GTC and 0.71 on the PTSD dataset. Notably, models trained on the GTC generalized well to other domains, suggesting that the explicit nature of tribunal testimonies provides a robust baseline for trauma detection.

GPT-4 also demonstrated strong zero-shot capabilities, nearly matching fine-tuned models on the PTSD dataset (F1 of 0.69). However, the Counseling dataset proved the most challenging for all models (GPT-4 F1 of 0.36), likely due to the subtle, emotional nature of the abuse described and significant class imbalance.

Discussion: The Universal Markers of Trauma

The XAI analysis revealed that while trauma is context-dependent, certain anchors like "death" and "sexual violence" serve as universal indicators. The model successfully distinguished between "war-related" trauma (e.g., bombardment, bullets) and "everyday" trauma (e.g., domestic abuse, accidents).

Crucially, the model passed the "Incel Control" test, ignoring misogynistic or aggressive language to focus specifically on traumatic events. This suggests the framework is not merely a "violence detector" but is sensitive to the specific narrative structures of trauma.

Conclusion and Future Directions

While the TRACE framework represents a significant step forward in cross-domain trauma modeling, limitations remain. Detecting mental and emotional abuse is still difficult compared to physical violence, as the language is often more subtle. Future work will focus on moving beyond binary classification to automatically categorize specific types of trauma and integrating these tools into digital platforms for real-time intervention and support.

Summary updated

BibTeX
@inproceedings{schirmer2024language,
  author = {Miriam Schirmer and Tobias Leemann and Gjergji Kasneci and Jürgen Pfeffer and David Jurgens},
  title = {{The Language of Trauma: Modeling Traumatic Event Descriptions Across Domains with Explainable AI}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.findings-emnlp.773},
}
Figure