LLMs-MentalHealth

This repository contains the materials related to the article "Addressing Class Imbalance and Explainability in the Detection of Mental Disorders from Social Media Data".

Requeriments

requirements.txt - required Python packages to run the code.

Dataset

For our experiments, we employed datasets from the Early risk prediction on the Internet (eRisk) forum, a widely recognized benchmark for mental health condition detection. Specifically, we selected datasets from the 2017, 2018, and 2020 editions, which correspond to depression, anorexia, and self-harm tendencies, respectively. All datasets were used in their original splits and are publicly available for research purposes.

Files

1. Preprocessing
2. Embeddings fastText, GloVe, Transformer-based models, and GPT
3. PBC4cip
4. LLMs GPT-4, LlaMA-3.1, FLANT-T5 and MENTAL-FLAN-T5

References

For more information, please read:

• CLEF 2017 eRisk Overview: Early Risk Prediction on the Internet: Experimental Foundations
• Overview of eRisk 2018: Early Risk Prediction on the Internet
• Overview of eRisk at CLEF 2020: Early Risk Prediction on the Internet