Go to ACL ARR 2024 December homepageFineThera: Fine-grained Therapeutic Alliance Prediction with Large Language Models
Abstract: Client perceptions of the therapeutic alliance are crucial predictors of counseling effectiveness, yet current methods of obtaining client feedback through questionnaires impose substantial burdens and are often impractical. This paper introduces a novel approach leveraging Large Language Models (LLMs) to automatically assess fine-grained dimensions of therapeutic relationships from counseling conversations.
We collect 3241 real-world counseling sessions and develop a comprehensive framework, utilizing 551 of these sessions with client-rated alliance scores across core therapeutic dimensions (goal, approach, bond) for training. Through rationale-augmented fine-tuning, our model not only outperforms human counselors in aligning with client perceptions (0.507 vs. 0.279 correlation) but also provides interpretable explanations for its predictions. Analysis of model-generated insights reveals key patterns in counselor behaviors that influence alliance formation, offering actionable guidance for improving therapeutic relationships.
Our work demonstrates the potential of LLMs to enhance counseling practice through automated, interpretable assessment while maintaining ethical considerations. The framework enables real-time understanding of client perspectives without additional burden, paving the way for more responsive and effective mental healthcare delivery.
Paper Type: Long
Research Area: Computational Social Science and Cultural Analytics
Research Area Keywords: NLP tools for social analysis, human behavior analysis
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis
Languages Studied: Chinese
Submission Number: 2360
Loading