Towards AI-Assisted Psychotherapy: Emotion-Guided Generative Interventions

Kilichbek Haydarov, Youssef Mohamed, Emilio Goldenhersch, Paul OCallaghan, Li-jia Li, Mohamed ElhoseinyEMNLP 2025 · OpenReview

TL;DR

This paper introduces a novel multimodal dataset of therapy sessions and a method for using non-verbal emotional cues (facial expressions and vocal tone) to improve AI-generated therapeutic interventions. Their approach, which incorporates "emotional dissonance" (mismatches between verbal and non-verbal cues), significantly enhances the empathy of AI responses, though it doesn't improve applicability or comprehensiveness.

What Problem Does This Paper Address?

Traditional AI approaches to mental health therapy primarily rely on text-based interactions, neglecting the crucial non-verbal emotional cues (like facial expressions and vocal tone) that are fundamental to human therapeutic practice. This limitation hinders the ability of AI to provide truly empathetic and contextually aware interventions. The authors aim to bridge this gap by developing methods that can effectively integrate these multimodal emotional signals into AI-driven therapeutic dialogue.

Methods

The core of the paper's methodology involves:

  • Multimodal Dataset Creation: They compiled a dataset of 1,441 publicly available, role-played therapy session videos. This dataset includes transcripts, speaker tags, timestamps, and crucially, non-verbal cues extracted from audio (speech emotion recognition - SER) and video (facial emotion recognition - FER).
  • Emotional Dissonance: Inspired by the concept of "emotional labor," they propose "emotional dissonance" as a computational measure of the mismatch between a patient's expressed emotion through words and their inferred emotion from non-verbal cues. This is formally defined as the maximum absolute difference between valence and arousal values from audio and video modalities, exceeding a certain threshold.
  • Emotion-Guided Prompting: They developed three prompting strategies for Large Language Models (LLMs):
  • Base Prompting: Uses only dialogue context.
  • Emotion-Enhanced Prompting: Augments prompts with categorical emotion labels or continuous valence-arousal values derived from non-verbal cues.
  • Dissonance-Based Prompting: Explicitly highlights instances of emotional dissonance in the prompt to draw the LLM's attention to these complex emotional signals.

Datasets

  • Newly Introduced Dataset: The authors created a novel multimodal dataset of 1,441 role-played therapy session videos. This dataset is unique in its combination of text, audio, and video modalities, specifically curated for mental health therapy research.
  • Comparison Datasets: The paper compares their dataset to existing mental health datasets like HOPE, MEMO, DAIC-WoZ, ALONE, Counsel-CHAT, and PAIR, highlighting the multimodal advantage of their contribution.

Key Insights and Ideas

  • Non-verbal cues are critical: The study strongly supports the hypothesis that integrating non-verbal emotional signals significantly enhances the empathy of AI-generated therapeutic interventions.
  • Emotional Dissonance is a valuable signal: The concept of emotional dissonance, representing a mismatch between verbal and non-verbal expressions, proves to be a powerful cue for LLMs to generate more nuanced and contextually aware responses.
  • Empathy is more sensitive to emotion: While empathy scores improved significantly with multimodal cues, applicability and comprehensiveness did not. This suggests that empathy is more directly influenced by emotional signals, whereas other aspects might rely more on contextual relevance and therapeutic framing.
  • LLM evaluators are not reliable for mental health: The study found a significant misalignment between LLM-based evaluations (using GPT-4o) and expert human judgments, particularly for applicability. This highlights a critical limitation of current automated evaluation methods in sensitive domains like mental health.
  • Expert evaluation is complex: Inter-rater reliability analysis revealed that differences in evaluators' training, experience, and therapeutic orientation can lead to varied assessments, underscoring the need for careful consideration of evaluator diversity and calibration.

Key Results

  • Empathy Improvement: All tested LLMs (GPT-4o, LLaMa3, Mistral-7b, Gemini) showed significant improvements in empathy ratings when using emotion-enhanced and dissonance-based prompts compared to the base prompt. For instance, GPT-4o and Gemini showed notable gains in empathy.
  • Applicability & Comprehensiveness: No significant improvements were observed for applicability and comprehensiveness criteria across the different prompting techniques.
  • Expert Evaluation: Five psychologists evaluated responses using a 1-to-5 Likert scale. Pairwise Cohen's kappa analysis showed substantial agreement for empathy and comprehensiveness, but moderate agreement for applicability.
  • LLM vs. Human Evaluation: Correlations between GPT-4o's ratings and human expert ratings were consistently low across all criteria, indicating a substantial misalignment.

Limitations and Future Work

  • LLM Hallucinations: The paper acknowledges that LLMs can generate hallucinated responses. Future work could explore Retrieval Augmented Generation (RAG) techniques.
  • Mock Therapy Data: While the dataset emulates real-world scenarios, it may not fully capture the complexities of actual therapy. The authors plan to incorporate real-world, clinician-reviewed interactions in future work.
  • Dataset Bias: The current dataset primarily features English-language, Western therapeutic frameworks and dyadic interactions, limiting generalizability across cultures, languages, or different therapeutic settings (e.g., group therapy). Future work should expand datasets to include diverse approaches and multilingual interactions.
  • LLM-Based Evaluation: The significant divergence between LLM and human evaluations highlights the need for more robust, domain-adapted automated metrics or hybrid human-AI evaluation approaches.
  • Ethical Considerations: The authors emphasize the risks of over-reliance on AI, misinformation, and the emotional impact of misinterpreting emotional cues. They advocate for human-in-the-loop evaluations and future work on explainability, bias detection, and adversarial robustness.

Summary updated

Abstract

Large language models (LLMs) hold promise for therapeutic interventions, yet most existing datasets rely solely on text, overlooking non-verbal emotional cues essential to real-world therapy. To address this, we introduce a multimodal dataset of 1,441 publicly sourced therapy session videos containing both dialogue and non-verbal signals such as facial expressions and vocal tone. Inspired by Hochschild’s concept of emotional labor, we propose a computational formulation of emotional dissonance—the mismatch between facial and vocal emotion—and use it to guide emotionally aware prompting. Our experiments show that integrating multimodal cues, especially dissonance, improves the quality of generated interventions. We also find that LLM-based evaluators misalign with expert assessments in this domain, highlighting the need for human-centered evaluation. Data and code will be released to support future research.
BibTeX
@inproceedings{haydarov2025towards,
  author = {Kilichbek Haydarov and Youssef Mohamed and Emilio Goldenhersch and Paul OCallaghan and Li-jia Li and Mohamed Elhoseiny},
  title = {{Towards AI-Assisted Psychotherapy: Emotion-Guided Generative Interventions}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=lynyujcOfi},
}
Figure