PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Interpretable Mental Health Analysis with Large Language Models

Kailai Yang, Shaoxiong Ji, Tianlin Zhang, Qianqian Xie, Ziyan Kuang, Sophia Ananiadou

arXiv:2304.03347v4cs.CL

TL;DR

Prior LLM studies of mental health analysis provided limited task coverage, prompting exploration, and explainability. This paper evaluates LLMs and emotional prompting strategies, finding that ChatGPT leads the tested LLMs but remains below advanced supervised methods, while its explanations approach human performance.

  • Problem

    Prior studies tested few binary detection tasks, lacked explanations, overlooked cause detection, and generally used prompts that ignored emotional cues.

    Source: §1 · p. 1

  • Method

    The paper evaluates LLMs across mental health tasks, compares emotional and few-shot prompting strategies, and assesses generated explanations with human ratings and automatic metrics.

    Source: §1 · p. 3, §2.2 · p. 4, §2.3 · p. 4

  • Results

    ChatGPT leads the tested LLMs but remains below advanced supervised methods; emotional prompting improves performance, and its explanations approach human level.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    The findings indicate potential for explainable mental health analysis and a need for customized automatic explanation-evaluation methods.

    Source: §5 · p. 9

    Rare ChatGPT refusals to classify were excluded from the experiments, leaving refusal behavior outside the evaluation.

    Source: Limitations · p. 10

Abstract

from arXiv · show

The latest large language models (LLMs) such as ChatGPT, exhibit strong capabilities in automated mental health analysis. However, existing relevant studies bear several limitations, including inadequate evaluations, lack of prompting strategies, and ignorance of exploring LLMs for explainability. To bridge these gaps, we comprehensively evaluate the mental health analysis and emotional reasoning ability of LLMs on 11 datasets across 5 tasks. We explore the effects of different prompting strategies with unsupervised and distantly supervised emotional information. Based on these prompts, we explore LLMs for interpretable mental health analysis by instructing them to generate explanations for each of their decisions. We convey strict human evaluations to assess the quality of the generated explanations, leading to a novel dataset with 163 human-assessed explanations. We benchmark existing automatic evaluation metrics on this dataset to guide future related works. According to the results, ChatGPT shows strong in-context learning ability but still has a significant gap with advanced task-specific methods. Careful prompt engineering with emotional cues and expert-written few-shot examples can also effectively improve performance on mental health analysis. In addition, ChatGPT generates explanations that approach human performance, showing its great potential in explainable mental health analysis.

1 Introduction

The paper addresses limited evaluation, prompting strategies, and explainability in prior LLM-based mental health analysis. It evaluates LLMs and prompting approaches, finding ChatGPT strongest among examined LLMs but below advanced supervised methods, while emotional cues and expert-written examples improve performance.

  • Prior studies focused on a few binary detection tasks, lacked explanations, and often overlooked cause detection and emotional cues.Source: §1 · p. 1
  • The study evaluates four LLMs across 11 datasets and five tasks, and creates a dataset of 163 human-assessed explanations.Source: §1 · p. 2
  • ChatGPT performs best among the examined LLMs but significantly underperforms advanced supervised methods on mental health analysis.Source: §1 · p. 2
  • Unsupervised emotion-enhanced CoT prompts perform best among the tested strategies, while expert-written few-shot examples also improve ChatGPT’s performance.Source: §1 · p. 2
  • ChatGPT generates explanations approaching human performance, while existing automatic metrics correlate only moderately with human judgments.The findings point to a need for customized automatic evaluation methods for explainable mental health analysis.Source: §1 · pp. 2–3

2 Methodology

The methodology benchmarks LLMs on mental health analysis, compares prompting approaches, and assesses generated explanations through human and automatic evaluation.

2.1 Large Language Models

  • The evaluated models include LLaMA-7B and LLaMA-13B, 13-billion-parameter InstructGPT-3, and ChatGPT.Source: §2.1 · p. 3

2.2 In-context Learning as Explainable Mental Health Analyzer

  • The analyzer compares zero-shot prompts, emotion-enhanced CoT, lexicon-derived emotion labels, and few-shot prompts with expert-written prediction-and-explanation examples.Source: §2.2 · pp. 3–4

2.3 Evaluation for Explainability

  • Human evaluators rate explanation fluency, reliability, completeness, and overall effectiveness on a 0–3 scale, with three annotators and domain-expert review.The evaluation includes 121 posts correctly classified by both models and 42 posts incorrectly classified by ChatGPT.Source: §2.3 · p. 4
  • Automatic evaluation benchmarks generation metrics including BLEU, ROUGE, GPT3-Score, BART-Score, and BERTScore-based methods against the annotated corpus.Source: §2.3 · pp. 4–5

3 Experimental Settings

The experimental settings cover mental health detection and cause identification across several datasets, using classification and explanation-evaluation metrics.

  • Datasets cover binary depression and stress detection, multi-class condition detection, and mental health cause or factor detection.The named datasets include Depression_Reddit, CLPsych15, Dreaddit, T-SID, SAD, and CAMS.Source: §3 · p. 5
  • Classification performance is measured with recall and weighted-F1, accounting for class imbalance and the importance of identifying false negatives.Source: §3 · p. 5
  • Explanation assessment uses Fleiss’ Kappa for evaluator agreement and Pearson correlation to compare automatic scores with human ratings.Source: §3 · p. 5

4 Results and Analysis

ChatGPT outperforms other evaluated LLMs on mental health analysis but remains behind fine-tuned methods; emotion-enhanced and expert few-shot prompts improve results. Human evaluations find strong explanations for correct classifications, while error analyses identify prompt sensitivity and weaknesses in reasoning on long posts.

4.1 Mental Health Analysis

  • Emotion-enhanced CoT beat other zero-shot prompts on most datasets, and expert few-shot examples improved T-SID 16.24%, SAD 6.88%, and CAMS 3.7%; lexicon sentiments often underperformed zero-shot.Fine-grained NRC EmoLex emotions beat NRC EmoLex sentiment on most datasets, including a 7.89% improvement on CAMS.Source: Table 1 · p. 6, §4.1 · p. 6
  • ChatGPT zero-shot significantly outperforms other LLMs across datasets but remains below fine-tuned methods, especially on T-SID.Zero-shot chain-of-thought prompting is comparable to or worse than plain zero-shot prompting.Source: §4.1 · pp. 5–6

4.2 Evaluation Results for Explainability

  • For correctly classified posts, ChatGPT explanations averaged over 2.5 overall, with reliability above 2.7; InstructGPT-3 averaged below 0.5 on completeness.ChatGPTtrue nearly averaged 3.0 in fluency; Figure 2 presents score distributions by aspect.Source: Figure 2 · p. 7, §4.2 · p. 7

4.3 Error Analysis

  • Incorrectly classified posts had comparable fluency but average completeness below 2.0; more than half had unreliable or inconsistent explanations below 1.0.These explanations averaged 335 tokens, over 130 more than those for correctly classified posts; the authors identify missed long-post information and unreliable reasoning as main error sources.Source: Figure 2 · p. 7, §4.3 · p. 9

4.2 Evaluation Results for Explainability

  • Over 95% of ChatGPT and 89.9% of InstructGPT-3 evaluations reached annotator agreement; all Fleiss’ Kappa values were ≥0.21 and 10 of 16 were ≥0.41.The authors interpret these results as evidence of human-annotation quality.Source: §4.2 · p. 7, Table 2 · p. 7
  • For ChatGPTtrue explanations, BART-Score had the highest human-correlation scores on every aspect and outperformed GPT3-Score in all aspects.The authors note that BART-Score’s pretraining on summarization and paraphrasing may benefit evaluation of relevance and coherence.Source: §4.2 · pp. 7–8, Table 3 · p. 8
  • For ChatGPTfalse explanations, BERT-Score performed best on fluency and reliability, while ROUGE led on completeness and overall scores.Mental-health-pretrained BERT-Score methods outperformed general BERT and RoBERTa methods on most aspects.Source: §4.2 · p. 8

4.3 Error Analysis

  • Adjective changes produced zero-shot variances of 10.6 on DR, 17.62 on CLPsych15, and 89.29 on Dreaddit, versus 1.34, 7.21, and 31.93 with few-shot examples.The best adjective differed by dataset; the authors attribute prompt sensitivity possibly to the subjective nature of mental health conditions.Source: Table 4 · p. 8, §4.3 · pp. 8–9

5 Conclusion

The paper evaluates LLMs for mental health analysis and explores explanations of their decisions. It finds that emotional information and prompt engineering can help, while ChatGPT still has limitations in reasoning and prediction stability.

5 Conclusion

  • The paper studies zero-shot and few-shot mental health analysis and the effects of emotion-enhanced prompts.It explores explaining predictions via Chain-of-Thought prompting.Source: §5 · p. 9
  • Emotional information with proper prompt engineering can better trigger LLMs’ mental health analysis ability, but mental health analysis remains challenging.The authors also report that ChatGPT has unstable predictions and inaccurate reasoning.Source: §5 · p. 9
  • ChatGPT can generate human-level explanations for its mental health analysis decisions, while current automatic evaluation metrics need further improvement.The work develops an expert-designed annotation protocol and benchmarks automatic metrics against human annotations.Source: §5 · p. 9
  • Future work will explore domain-specific LLM fine-tuning to address inaccurate reasoning and extend interpretable settings to other research domains.Source: §5 · p. 9

Limitations

The limitations include refusals to classify in a small number of cases and constraints in the sentiment and emotion lexicons used for prompting.

Limitations

  • A few ChatGPT responses refuse classification because a post lacks evidence or violates OpenAI’s content policy; these rare responses were excluded from the experiments.The authors note that future efforts are needed to alleviate the issue.Source: Limitations · p. 10
  • The decade-old VADER and NRC EmoLex lexicons may carry annotation bias, reflect older language, and have limited vocabularies.The authors also note that manual rules for aggregating sentence-level sentiment and emotions may be underspecified, and that ChatGPT’s black-box nature prevents clear attribution of VADER’s slight T-SID improvement.Source: Limitations · p. 10

Ethical Considerations

The paper frames social-media mental health analysis as a nonclinical aid and describes safeguards for handling sensitive posts. Its related-work discussion reviews PLM and LLM approaches while noting that prior LLM evaluations cover relatively narrow task sets.

Ethical Considerations

  • The study uses public Reddit and Twitter posts under privacy protocols, disguises paper examples, and positions predictions as nonclinical support rather than psychiatric diagnoses.The authors note that post intentions may not be understood and interpretations can differ because some disorders are subjective.Source: Ethical Considerations · p. 10

A.1 Mental Health Condition Detection

  • Prior mental health detection work includes BERT-based classifiers, social-media PLMs, metaphor features, and grounding predictions in PHQ9 symptoms.The related work also describes multimodal approaches that combine text with speech or visual information for depression detection.Source: §A.1 · p. 14

A.2 Large Language Models for Mental Health Analysis

  • The reviewed LLM studies include mental health classification and emotional reasoning, but most cover only simple sentiment or a few binary mental-health tasks.The paper identifies this as a gap in comprehensive evaluation of emotion-aware mental health analysis.Source: §A.2 · pp. 14–15

B ChatGPT for Emotional Reasoning

ChatGPT’s zero-shot emotional reasoning performs competitively against some lighter supervised methods but remains behind advanced task-specific systems. The section evaluates emotion recognition in conversations and causal emotion entailment using direct zero-shot prompts.

B ChatGPT for Emotional Reasoning

  • ERC assigns an emotion label to each utterance from a fixed category set, while CEE identifies candidate utterances that cause a target emotion in prior dialogue.The evaluation uses four ERC datasets and RECCON for CEE, with task-specific F1 metrics.Source: B ChatGPT for Emotional Reasoning · p. 15
  • The study compares direct zero-shot classification prompts against supervised baselines for both emotional reasoning tasks.Source: B ChatGPT for Emotional Reasoning · p. 15
  • 61.18% weighted-F1 on MELD: ChatGPTZS exceeds fine-tuned BERTBase by 4.97% and KET by 3.0%, though it remains below advanced supervised methods overall.Source: B ChatGPT for Emotional Reasoning · p. 16, Table 5 · p. 16
  • 19.86% macro-F1 gap to KBCIN: ChatGPTZS also outperforms RankCP on positive F1, but trails advanced supervised methods on all metrics.The authors identify positive F1 as more indicative of emotion-cause detection ability than RankCP’s highest negative F1.Source: B ChatGPT for Emotional Reasoning · p. 16, Table 6 · p. 16
  • Across ERC and CEE, ChatGPT is comparable to some traditional methods but strongly underperforms competitive task-specific information-infusion and fine-tuning methods.The authors point to prompting strategies and external knowledge as directions for strengthening emotional reasoning.Source: B ChatGPT for Emotional Reasoning · p. 16

C.1 Emotional Reasoning

  • The ERC prompt supplies dialogue history, the target utterance, and the dataset’s predefined emotion list, and requests only the assigned label.Source: §C.1 · p. 16
  • The CEE prompt supplies dialogue history with emotion labels, a candidate utterance, and the target emotion, and requests a Yes/No causal judgment.Source: §C.1 · p. 16

C.2 Mental Health Analysis

The mental-health analysis setup varies prompt form by model, adds emotional cues through prompting, and covers emotion-recognition and causal-entailment baselines. These choices span several task types and ways of representing conversational context.

C.2 Mental Health Analysis

  • LLaMA uses completion-based prompts because it is not instruction-tuned, while InstructGPT-3 and ChatGPT use instruction-based prompts for classification.Source: §C.2 · p. 17
  • Emotion-enhanced CoT prompts ask ChatGPT to consider post emotions and explain its reasoning step by step for detection and cause/factor tasks.Source: §C.2 · p. 17
  • Supervised emotion-enhanced prompts inject sentiment or emotion labels assigned to posts using sentiment and emotion lexicons.Source: §C.2 · p. 17
  • Few-shot examples pair task instructions with worked responses and stepwise reasoning for mental-health condition and stress-cause detection.Source: §C.2 · pp. 17–18

D.1 Emotion Reasoning

  • Emotion-reasoning baselines span context-free utterance classification and models that represent conversational context or long-range dependencies.Source: §D.1 · p. 18
  • Causal emotion entailment baselines include clause-pair ranking, utterance-pair classification, and models incorporating social commonsense knowledge.Source: §D.1 · p. 19

D.2 Mental Health Analysis

The section describes mental-health baselines and human-rating criteria, then illustrates emotion-enhanced prompting with a stress-detection example. The example combines emotional and physical evidence in its prediction rationale.

D.2 Mental Health Analysis

  • Mental-health baselines range from CNN and recurrent models to BERT/RoBERTa and domain-specific MentalBERT/MentalRoBERTa classifiers.Source: §D.2 · pp. 19–20

E Human Evaluation Criteria

F.1 Emotion-enhanced Prompts

  • In the Dreaddit example, ChatGPT identifies fear, anxiety, and distress and considers emotional and physical symptoms; the authors say emotion-enhanced zero-shot CoT improves overall performance.Source: §F.1 · p. 21

F.2 Explainability

The examples show ChatGPT explaining binary, multiclass, and cause-detection decisions with post-specific evidence. A separate case illustrates an error when it attributes quoted distress to the poster.

F.2 Explainability

  • For binary stress detection, ChatGPT cites fear-related language and prolonged adrenaline use as emotional and physical symptoms supporting its prediction.Source: §F.2 · p. 21
  • In the multiclass example, ChatGPT labels the post Suicide, citing explicit suicidal ideation and explaining why depression and PTSD are not established.Source: §F.2 · p. 21
  • For depression-cause detection, ChatGPT selects Jobs and career from the firing detail and connects the complaint with possible anxiety.Source: §F.2 · pp. 21–22

F.3 Inaccurate Reasoning

  • In the inaccurate stress prediction, ChatGPT mistakes another girl's quoted relationship problem for the poster's own circumstances.Source: §F.3 · p. 22

G Automatic Evaluation on InstructGPT-3

The automatic-metric comparison reports BERT-Score as strongest across aspects for InstructGPT-3 explanations. The authors note that many outputs provide diagnoses without explanations.

Loading 2304.03347v4…