Fixaro/myanmar-absa-sentiment-classification

Myanmar ABSA - Sentiment Classification Model (Stage 2)

An aspect-level sentiment classification model for Burmese product/service reviews. This is Stage 2 of a two-stage Aspect-Based Sentiment Analysis (ABSA) pipeline that predicts sentiment polarity given a review and an aspect category.

Model Details

• Base Model: xlm-roberta-base
• Architecture: XLM-RoBERTa for Sequence Classification
• Problem Type: Single-label classification (sentence-pair)
• Language: Burmese (Myanmar)
• License: MIT

Sentiment Classes

The model predicts sentiment polarity for a given aspect:

• 0: negative - Negative sentiment or dissatisfaction
• 1: neutral - Neutral or factual statements
• 2: positive - Positive sentiment or satisfaction

Input Format

This model uses sentence-pair classification:

• Text: The review text
• Text Pair: The aspect category (one of 5 aspects)

Supported Aspects

1. product_quality
2. fulfillment_and_speed
3. price_and_value
4. staff_and_service
5. variety_and_availability

Training Data

• Total Samples: 5,247 aspect-level annotations
• Data Split: Stratified train/val/test split
  • Train: 4,197 samples
  • Validation: 525 samples
  • Test: 525 samples
• Alignment: Filtered to match Stage 1 aspect taxonomy
• Label Distribution:
  • Negative: 2,107 samples (40.2%)
  • Neutral: 406 samples (7.7%)
  • Positive: 2,734 samples (52.1%)

Performance Metrics

Evaluated on the test set:

Metric	Score
F1 Macro	0.8671
F1 Micro	0.9124
F1 Weighted	0.9115

Per-Class Performance

Class	Precision	Recall	F1-Score
negative	0.92	0.92	0.92
neutral	0.81	0.71	0.75
positive	0.92	0.94	0.93

Training Configuration

• Learning Rate: 2e-5
• Batch Size: 8 (with gradient accumulation steps = 4, effective batch size = 32)
• Epochs: 4
• Max Sequence Length: 128 tokens
• Optimizer: AdamW
• Precision: bfloat16

Usage

Basic Usage with Pipeline

from transformers import pipeline

# Load the model
classifier = pipeline(
    "text-classification",
    model="Fixaro/myanmar-absa-sentiment-classification"
)

# Predict sentiment for a specific aspect
review = "ပစ္စည်းအရည်အသွေးက ကောင်းပါတယ်"
aspect = "product_quality"

result = classifier({
    "text": review,
    "text_pair": aspect.replace('_', ' ')  # Convert to readable format
})

print(f"Sentiment: {result[0]['label']}")
print(f"Confidence: {result[0]['score']:.4f}")

Advanced Usage with Tokenizer

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "Fixaro/myanmar-absa-sentiment-classification"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Prepare input (sentence pair)
review = "ဈေးလည်း သက်သာတယ်၊ ပို့တာလည်း မြန်တယ်"
aspect = "price_and_value"

inputs = tokenizer(
    review,
    aspect.replace('_', ' '),  # Convert to readable format
    return_tensors="pt",
    truncation=True,
    max_length=128,
    padding=True
)

# Get predictions
with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits
    probabilities = torch.softmax(logits, dim=-1)
    predicted_class = torch.argmax(probabilities, dim=-1).item()

# Map to sentiment label
sentiment_labels = {0: "negative", 1: "neutral", 2: "positive"}
predicted_sentiment = sentiment_labels[predicted_class]

print(f"Aspect: {aspect}")
print(f"Sentiment: {predicted_sentiment}")
print(f"Confidence: {probabilities[0][predicted_class]:.4f}")

Complete ABSA Pipeline Example

from transformers import pipeline

# Load both models
aspect_detector = pipeline(
    "text-classification",
    model="Fixaro/myanmar-absa-aspect-detection",
    return_all_scores=True,
    function_to_apply="sigmoid"
)

sentiment_classifier = pipeline(
    "text-classification",
    model="Fixaro/myanmar-absa-sentiment-classification"
)

# Analyze a review
review = "ပစ္စည်းက ကောင်းတယ်၊ ဒါပေမယ့် ပို့တာ နည်းနည်း ကြာတယ်"

# Stage 1: Detect aspects
aspect_scores = aspect_detector(review)[0]
detected_aspects = [
    result['label'] 
    for result in aspect_scores 
    if result['score'] > 0.5
]

# Stage 2: Classify sentiment for each aspect
results = []
for aspect in detected_aspects:
    sentiment_result = sentiment_classifier({
        "text": review,
        "text_pair": aspect.replace('_', ' ')
    })
    results.append({
        "aspect": aspect,
        "sentiment": sentiment_result[0]['label'],
        "confidence": sentiment_result[0]['score']
    })

# Display results
print(f"Review: {review}\n")
for result in results:
    print(f"Aspect: {result['aspect']}")
    print(f"Sentiment: {result['sentiment']} ({result['confidence']:.4f})\n")

Model Architecture

XLM-RoBERTa Base (xlm-roberta-base)
├── Encoder: 12 transformer layers
├── Hidden size: 768
├── Attention heads: 12
└── Classification head: Linear(768, 3) with softmax activation

Limitations and Biases

• Aspect Dependency: Performance depends on accurate aspect detection from Stage 1
• Class Imbalance: Neutral class has fewer samples (7.7%), leading to lower recall
• Context Sensitivity: May struggle with sarcastic or context-dependent sentiment
• Aspect Specificity: Requires explicit aspect input; cannot infer aspect from context
• Domain Specificity: Trained on product/service reviews; may not generalize to other domains

Intended Use

This model is intended for:

• Aspect-level sentiment analysis in Burmese reviews
• Second stage of ABSA pipelines (after aspect detection)
• Fine-grained sentiment analysis
• Research and academic purposes
• Commercial applications with proper validation

Citation

If you use this model in your research, please cite:

@software{myanmar_absa_sentiment_classification,
  title = {Myanmar ABSA: Sentiment Classification Model},
  author = {Fixaro},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/Fixaro/myanmar-absa-sentiment-classification}
}

Related Models

• Stage 1 Model: Fixaro/myanmar-absa-aspect-detection - Multi-label aspect detection

Contact

For questions or issues, please open an issue on the model repository.