ChatGPT for Suicide Risk Assessment on Social Media: Quantitative Evaluation of Model Performance, Potentials and Limitations

Hamideh Ghanadian, Isar Nejadgholi, Hussein Al OsmanACL 2023

TL;DR

This paper evaluates ChatGPT's ability to assess suicide risk from social media posts, comparing it to fine-tuned transformer models. While ChatGPT shows promise, fine-tuned models like ALBERT achieve superior performance, highlighting the need for careful hyperparameter tuning and expert-in-the-loop systems for sensitive applications.

What Problem Does This Paper Address?

Suicide remains a critical global health issue, with social media platforms offering a potential avenue for early identification and intervention. However, accurately assessing suicide risk from the vast and often nuanced text data on these platforms is challenging. Existing Natural Language Processing (NLP) techniques have been applied, but the emergence of powerful large language models (LLMs) like ChatGPT necessitates a thorough evaluation of their capabilities and limitations in this sensitive domain. This paper specifically investigates how effectively ChatGPT can assess suicide risk from social media posts, comparing its performance to established methods and exploring the impact of its internal parameters.

Methods

The researchers employed a multi-faceted approach to evaluate ChatGPT's suicide risk assessment capabilities:

  • Zero-Shot Learning: ChatGPT was tasked with assessing suicide risk without any prior specific training on suicide-related data. This leverages its general language understanding to interpret the provided text based on predefined risk criteria.
  • Few-Shot Learning: In this setting, ChatGPT was provided with a small number of labeled examples (e.g., 8 examples) to guide its assessment. This approach aims to see if a limited amount of context can significantly improve its performance.
  • Hyperparameter Tuning (Temperature): The study systematically varied the "temperature" parameter in ChatGPT. This parameter controls the randomness of the model's output; lower temperatures lead to more deterministic and focused responses, while higher temperatures encourage more diverse and creative (but potentially less accurate) outputs.
  • Comparison with Fine-Tuned Models: The performance of ChatGPT was benchmarked against two established transformer-based models, ALBERT and DistilBERT, which were fine-tuned on the specific suicide risk assessment task using labeled data.

Datasets

The primary dataset used in this study is the University of Maryland Reddit Suicidality Dataset (UMD). This dataset comprises over 100,000 Reddit posts and comments from individuals expressing suicidal thoughts or behaviors, collected from subreddits like "r/SuicideWatch." For this research, a specific subset (Task A) was utilized, focusing on user-level risk assessment. This subset was further divided into labeled and unlabeled users, and then split into training and testing sets.

Key Insights and Ideas

  • ChatGPT's Potential, But Not Superiority: While ChatGPT demonstrates a notable ability to assess suicide risk, it does not outperform fine-tuned transformer models like ALBERT. This suggests that specialized training on domain-specific data remains crucial for achieving peak performance in this critical area.
  • Temperature's Double-Edged Sword: The temperature parameter significantly impacts ChatGPT's performance. Lower temperatures lead to higher accuracy on instances where a decision is made, but also increase the "inconclusiveness rate" (i.e., the model stating it cannot provide an assessment). Conversely, higher temperatures reduce inconclusiveness but can lead to more errors.
  • Few-Shot Learning Boosts Confidence: Providing ChatGPT with even a few examples (Few-Shot Learning) significantly reduces its inconclusiveness rate, indicating it becomes more confident in its assessments. This suggests that a hybrid approach, combining LLM capabilities with targeted examples, could be highly effective.
  • The Challenge of Subjectivity: The study highlights that the "Low Risk" and "Moderate Risk" categories are particularly challenging for ChatGPT, likely due to their inherent subjectivity and less clear-cut definitions. This underscores the difficulty of automating nuanced human judgment.
  • Expert-in-the-Loop is Key: The findings strongly support the idea of using ChatGPT as an assistive tool for human annotators or mental health professionals, rather than a fully autonomous system. This "expert-in-the-loop" approach can speed up annotation and improve accuracy while mitigating risks.

Key Results

  • Zero-Shot Learning: ChatGPT achieved an F1-score of 0.73 at a temperature of 0.1. The inconclusiveness rate was highest at this temperature (2.91%) and decreased as temperature increased, while the F1-score also decreased.
  • Few-Shot Learning: ChatGPT achieved an F1-score of 0.71 at a temperature of 0.1. The inconclusiveness rate was significantly lower (0.58%) and remained relatively constant across different temperatures, suggesting increased confidence.
  • Comparison with Fine-Tuned Models:
    • Fine-tuned ALBERT achieved an F1-score of 0.869.
  • Fine-tuned DistilBERT achieved an F1-score of 0.745.
  • ChatGPT (Zero-Shot, temp=0.1) achieved an F1-score of 0.73.
  • ChatGPT (Few-Shot, temp=0.1) achieved an F1-score of 0.71. This shows that fine-tuned ALBERT significantly outperformed both DistilBERT and ChatGPT.

Limitations and Future Work

The authors acknowledge several limitations:

  • Small Test Dataset: The study was conducted on a relatively small test dataset, which may limit the generalizability of the findings.
  • Unknown Training Data: The specifics of ChatGPT's training data are not disclosed, making it impossible to definitively rule out its prior exposure to the UMD dataset.
  • Focus on Initial Assessment: The study focused solely on the initial step of suicide risk assessment and did not explore ongoing monitoring or intervention.
  • Absence of Other Safety Criteria: The evaluation did not measure other critical safety aspects like biases or privacy issues.

Future work should focus on:

  • Larger and Diverse Datasets: Evaluating ChatGPT on larger, more diverse datasets to assess generalizability.
  • Bias Mitigation: Developing techniques to mitigate biases present in LLMs.
  • Other Chatbots: Evaluating open-source conversational models and advanced chatbots like Vicuna.
  • Expert-in-the-Loop Systems: Further exploring and refining ChatGPT's role in expert-in-the-loop systems for suicide risk assessment.
  • Broader Clinical Settings: Investigating the feasibility, safety, and effectiveness of ChatGPT in broader clinical applications.

Summary updated

BibTeX
@inproceedings{ghanadian2023chatgpt,
  author = {Hamideh Ghanadian and Isar Nejadgholi and Hussein Al Osman},
  title = {{ChatGPT for Suicide Risk Assessment on Social Media: Quantitative Evaluation of Model Performance, Potentials and Limitations}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure