---
license: cc-by-nc-4.0
language:
- fr
base_model:
- EuroBERT/EuroBERT-210m
library_name: transformers
pipeline_tag: text-classification
tags:
- sentiment 
- french
- eurobert
- classification
- text
- bert
---

# EuroBERT-Sentiment-Analysis-french

## Usage

```python
from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="Ant-Dlc/EuroBERT-SenimentAnalysis-French",
    truncation=True,
    trust_remote_code=True,
    
)

clf("[text]")
```

## Model description
**EuroBERT-Sentiment-Analysis-french** is a sentiment analysis model for French verbatims answering to the question "Qualification équipement" from an IT satisfaction survey.  
It classifies a given text into three categories:  
- **NEGATIF**  (french for "negative")
- **NEUTRE**   (french for "neutral")
- **POSITIF**  (french for "positive")

The model is based on **EuroBERT (210M parameters)** and has been fine-tuned using **Hugging Face Transformers** and **PyTorch**.  

---

## Training data
The model was trained on approximately 50,000 verbatims from an IT satisfaction survey. Labels were generated using an LLM (Mistral Small 24B, see Ant-Dlc/LLMPrompt-SentimentAnalysis)

⚠️ Notes on the dataset:  
- The dataset is not perfectly balanced between classes.  
- Duplicates were kept when they appeared in the survey (e.g., common responses such as *"basique"* or *"RAS"*).  

---

## Evaluation
For evaluation, **500 verbatims were manually labeled** and compared with the model predictions.  

- To avoid data leakage, any test sample that also appeared in the training set was adjusted.  
  - Example: if the training set contained *10 times* the word "basique" and the test set contained *3 times* the same word, then *7 occurrences* were kept in the training set.
  - We made this choice because the presence of multiple occurrences of the same word reflects the nature of the problem we aim to solve.
    
The confusion matrix below summarizes the results (to be inserted as an image or table):  

![Confusion Matrix](confusion_matrix.png)

## Observations

- About half of the errors are **subjective** (e.g., the word *"moyen"* classified as **NEUTRAL** instead of **NEGATIVE**).  
- The model sometimes **overweights local negatives** in mixed statements where the overall verdict is positive but followed by specific drawbacks (e.g., *"globalement c’est bien, mais …"*). These should be **POSITIVE** but are often predicted **NEGATIVE**.
- A few rare **misclassifications** occur (**POSITIVE** instead of **NEGATIVE**), but they represent only a small fraction of errors.  

Overall, performance is good and consistent with human judgment, though some ambiguity in labels remains.