Burmese Sentiment Analysis with XLM-RoBERTa
Model Details
Model Description
This model is a fine-tuned version of FacebookAI/xlm-roberta-base for Burmese sentiment analysis.
It classifies Burmese text into one of three sentiment categories:
- Positive
- Negative
- Neutral
The model was trained using publicly available Burmese sentiment datasets and additional manually curated data, with careful preprocessing to normalize encoding (Zawgyi → Unicode conversion).
- Developer: Yoon Thiri Aung (GitHub)
- Model type: Transformer-based multilingual masked language model fine-tuned for text classification
- Languages: Burmese (
my), with multilingual base model support - License: MIT
- Finetuned from: FacebookAI/xlm-roberta-base
- Demo: https://huggingface.co/spaces/emilyyy04/burmese-sentiment-analysis-demo
Uses
Direct Use
- Sentiment classification of Burmese text from social media, reviews, comments, and other user-generated content.
- Building sentiment-aware Burmese NLP applications such as chatbots, analytics dashboards, and content moderation tools.
Limitations
- May not generalize well to domains significantly different from the training data.
- May misclassify sentences with mixed sentiments or sarcasm.
- Performance may drop for code-mixed Burmese-English text with heavy slang or informal spelling.
Training Details
Training Data
-
Sources:
kalixlouiis/burmese-sentiment-analysischuuhtetnaing/myanmar-social-media-sentiment-analysis-dataset- Additional curated data collected and annotated by the author.
-
Preprocessing:
- Converted Zawgyi-encoded text to Unicode.
- Cleaned and normalized text fields.
- Tokenized using the XLM-RoBERTa tokenizer with:
max_length=128- Truncation and padding to maximum length.
Training Procedure
- Optimizer: AdamW (default in Hugging Face
Trainer) - Learning rate: 2e-5
- Batch size: 8 (train & eval)
- Epochs: 3
- Weight decay: 0.01
- Mixed precision (fp16): Enabled when training on GPU
- Metric for best model: F1 score (weighted average)
- Evaluation strategy: Per epoch
- Model selection: Best F1 score checkpoint
Evaluation
Metrics
The model was evaluated on a held-out validation set using accuracy, precision, recall, and F1 score.
| Epoch | Val Loss | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|---|
| 1 | 0.6171 | 0.7859 | 0.7994 | 0.7859 | 0.7875 |
| 2 | 0.4268 | 0.8470 | 0.8465 | 0.8470 | 0.8464 |
| 3 | 0.4115 | 0.8451 | 0.8447 | 0.8451 | 0.8448 |
The final model used is the checkpoint with the highest F1 score.
How to Use
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_name = "emilyyy04/burmese-sentiment-xlm-roberta" # Replace with actual repo name
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
text = "ဒီဇာတ်လမ်းက တကယ်ကောင်းတယ်။"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
outputs = model(**inputs)
predicted_class = torch.argmax(outputs.logits, dim=1).item()
label_map = {0: "positive", 1: "negative", 2: "neutral"}
print("Predicted Sentiment:", label_map[predicted_class])