---
license: mit
language:
  - en
  - fa
tags:
  - therapy
pretty_name: Multilingual Therapy Dialogues
size_categories:
  - 1K<n<10K
---

[![arXiv](https://img.shields.io/badge/arXiv-Paper-<COLOR>.svg)]() [![GitHub](https://img.shields.io/badge/GitHub-Code-181717?logo=github)]()

## Dataset Summary
**Multilingual Therapy Dialogues** is a diverse and bilingual dataset consisting of paired dialogues between patients and therapists in both Persian and English.

## Dataset Statistics
- Number of samples: 7,179  

**English:**
1. Average tokens per sentence: 101.30  
2. Maximum tokens in a sentence: 939  
3. Average characters per sentence: 567.85  
4. Number of unique tokens: 32,968  

**Persian:**
1. Average tokens per sentence: 100.06  
2. Maximum tokens in a sentence: 1,413  
3. Average characters per sentence: 516.57  
4. Number of unique tokens: 33,298  

## Dataset Fields
1. **Patient**: Original English text spoken by the patient.  
2. **Therapist**: Original English text spoken by the therapist.  
3. **Translated Patient**: Persian translation of the patient's text.  
4. **Translated Therapist**: Persian translation of the therapist's text.  

## Dataset Generation Pipeline

The dataset was constructed using the following steps:

1. **Data Collection**: Dialogues were collected from various public sources, including:
   - [Mental Health Counseling Conversations](https://huggingface.co/datasets/Amod/mental_health_counseling_conversations)
   - [Mental Health CSV Dataset](https://www.kaggle.com/datasets/zuhairhasanshaik/datacsv)
   - [Mental Health Conversational Data](https://www.kaggle.com/datasets/elvis23/mental-health-conversational-data)
   - Additional manually curated sources

2. **Translation**: English dialogues were translated into Persian using the [SeamlessM4T model](https://github.com/facebookresearch/seamless_communication) by Meta AI.

3. **Refinement**: Translations were revised and enhanced in three steps using GPT-4o:
   - First pass to make the tone more natural and emotionally sympathetic to be more likely to real world scenarios.
   - Second pass to improve fluency and human-likeness.
   - Final pass for consistency and correction of subtle translation errors.

4. **Filtering**: Only meaningful and conte


## Usage Instructions

### Option 1: Manual Download

Visit the [dataset repository](https://huggingface.co/datasets/Algorithmic-Human-Development-Group/Multilingual-Therapy-Dialogues/tree/main) and download the `SAT_dataset.csv` file.

### Option 2: Programmatic Download

Use the `huggingface_hub` library to download the dataset programmatically:

```python
from huggingface_hub import hf_hub_download
import pandas as pd

dataset = hf_hub_download(
    repo_id="Algorithmic-Human-Development-Group/Multilingual-Therapy-Dialogues",
    filename="SAT_dataset.csv",
    repo_type="dataset"
)
df = pd.read_csv(dataset)
df.head()
```

## Citations
If you find our paper, code, data, or models useful, please cite the paper:  
```
To be updated once the paper is published.
```

## Contact
If you have questions, please email sinaaelahimanesh@gmail.com or mahdi.abootorabi2@gmail.com.