LoCCA-Frame

Official implementation of LoCCA-Frame, a mental health counseling framework presented at KCC 2025.

This project combines MentalRoBERTa-based disorder classification with GPT-4o response generation to provide strategy-driven counseling.

Paper

Title: LoCCA-Frame: A Low-Cost and Strategy-Guided Prompting Framework for Enhancing Diagnosis-Prediction Mental Health Counseling (LoCCA-Frame: 질환 예측 기반 정신 건강 상담 역량 향상을 위한 저비용 전략 프롬프트 프레임워크)

Conference: Korea Computer Congress 2025 (KCC 2025)

Abstract

Generic LLMs often hallucinate or provide ungrounded advice in mental health contexts. This framework addresses these issues using a two-stage pipeline:

1. Classification: A fine-tuned MentalRoBERTa model identifies the user's specific mental disorder (e.g., Depression, Anxiety).
2. Generation: MentaLLaMA generates a response conditioned on clinically verified treatment strategies associated with the detected disorder.

Project Structure

LoCCA-Frame/
├── src/
|   |── eval/                       # Evaluation Metric
│   ├── MentalRoBERTa_Training.py   # Fine-tuning script for classification
│   ├── pipeline.py                 # Main inference pipeline
│   └── Trained_MentalRoBERTa/      # Saved model checkpoints
├── data/
│   ├── strategy_info.json          # Mapping of disorders to treatment 
└── README.md

Usage

1. Classification Model Training

Fine-tune MentalRoBERTa on the mental health dataset.

python src/MentalRoBERTa_Training.py

The model will be saved to ./src/Trained_MentalRoBERTa/best_model/.

2. Inference Pipeline

Run the full pipeline (Classification + Strategy Retrieval + Generation).

python src/pipeline.py --data_file ./data/your_data.jsonl

Arguments:

• --api_key: (Required) OpenAI API key.
• --data_file: Path to input JSONL file.
• --output_file: Name of the output file (default: total_inference.jsonl).

Data

The model was trained on a combined dataset of SWMH, Reddit Mental Health Diagnoses, and Comprehensive PTSD Analysis Dataset.

Classes: Depression, Anxiety, PTSD, Bipolar Disorder, Eating Disorder.

Note: Raw datasets are not included in this repository due to licensing/privacy restrictions.

Citation

@inproceedings{Kim2025LoCCA,
  title={LoCCA-Frame: 질환 예측 기반 정신 건강 상담 역량 향상을 위한 저비용 전략 프롬프트 프레임워크 (LoCCA-Frame: Low-Cost Strategic Prompt Framework for Improving Mental Health Counseling Competence Based on Disorder Prediction)},
  author={Kim, Kwangil and Kim, Seulgi and Park, Seongbae and Yoon, Taewoong and Song, Jiwoo and Choi, Yeonseok and Kim, Harksoo},
  booktitle={Proceedings of the Korea Computer Congress 2025 (2025 한국컴퓨터종합학술대회 논문집)},
  pages={566--568},
  year={2025},
  organization={Korean Institute of Information Scientists and Engineers (한국정보과학회)}
}