Classificador de Sentiments a Xarxes Socials en Català (CSXSC)

This model is a fine-tuned version of projecte-aina/roberta-base-ca-2 for three-class sentiment analysis (positive, negative, neutral) of Catalan-language texts from social media and reviews.

Developed as part of a Master's Thesis, the CSXSC model achieves high performance on sentiment classification while being exceptionally efficient, making it ideal for production environments. It was trained on the custom-built Danie1Arias/sentiment-analysis-catalan-reviews dataset, which was specifically created and balanced for this task.

---

Intended Uses

This model is designed for analyzing the sentiment of user-generated content in Catalan. It's particularly well-suited for:

• Customer Feedback Analysis: Automatically classify customer reviews on products or services.
• Social Media Monitoring: Track public opinion and brand sentiment on social platforms.
• Opinion Mining: Analyze sentiment in forum posts and comments.

---

Model Details

• Architecture: RoBERTa (encoder-only)
• Base Model: projecte-aina/roberta-base-ca-v2
• Language: Catalan (ca)
• Fine-tuning Task: Sequence Classification
• Labels: negative, neutral, positive

---

Training Procedure

The model was fine-tuned for 5 epochs using the AdamW optimizer with a learning rate of 2e-5, a batch size of 32, and a weight decay of 0.01. A linear learning rate scheduler with a warmup phase over the first 10% of training steps was employed. The final model selected was the one that achieved the highest Quadratic Weighted Kappa (QWK) score on the validation set, ensuring optimal generalization.

---

Evaluation

The model was evaluated on a held-out test set, achieving the following results:

Metric	Score
Accuracy	83.69%
F1 Macro	0.8186
Quadratic Weighted Kappa (QWK)	0.8715

---

Limitations and Bias

• This model is specialized for general sentiment in social media and reviews. Its performance may degrade on highly specialized or technical domains (e.g., legal or medical texts).
• Like all language models, it may struggle with sarcasm, irony, or complex figurative language.
• The training data is sourced from public forums; therefore, the model may reflect biases present in that data.

---

Usage Example

Here’s how to load and use the model to get human-readable predictions:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the model and tokenizer
model_name = "Danie1Arias/CSXSC"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Input text
text = "El servei ha estat una mica lent, però el menjar era boníssim." 
# -> "The service was a bit slow, but the food was delicious."

# Tokenize and predict
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
predicted_class_id = torch.argmax(outputs.logits, dim=1).item()

# Map the prediction ID to the label name
predicted_label = model.config.id2label[predicted_class_id]

print(f"Text: '{text}'")
print(f"Predicted Label: {predicted_label}")
# Expected output: Predicted Label: positive

---

Citation & Repository

This model was developed as part of a Master's Thesis. If you use this model or its dataset in your research, please cite the original work.

The full code for data processing, training, and evaluation is available on GitHub: https://github.com/Danie1Arias/CSXSC/

@mastersthesis{arias2025csxsc,
  title={From Traditional to Large Language Models: A Novel NLP-Based Model for Sentiment Analysis in Social Media},
  author={Arias Cámara, Daniel},
  year={2025},
  school={Rovira i Virgili University}
}