Fixaro/myanmar-absa-sentiment-classification

Myanmar ABSA - Stage 2 (Aspect-Level Sentiment Classification)

This model is a fine-tuned version of xlm-roberta-base trained on Burmese customer reviews formatted as sentence pairs (review_text, target_aspect). It predicts the sentiment polarity for a given target aspect.

Sentiment Classes:

• 0: Negative
• 1: Neutral
• 2: Positive

---

Supported Aspect Names:

1. product_or_service_quality
2. fulfillment_and_speed
3. price_and_value
4. digital_experience
5. customer_support
6. variety_and_availability

---

Usage Example (Sentence-Pair Classification)

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

MODEL_NAME = "Fixaro/myanmar-absa-sentiment-classification"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)

# Input text and the specific aspect to evaluate
review_text = "ဒီဆိုင်က ဝန်ဆောင်မှု အရမ်းကောင်းတယ်၊ ဒါပေမဲ့ ပို့ဆောင်ခ အရမ်းကြီးတယ်။"
target_aspect = "customer_support"

# 1. Tokenize as a Sentence Pair (Review Text + Target Aspect)
inputs = tokenizer(review_text, target_aspect, return_tensors="pt", truncation=True, max_length=128)

# 2. Predict Sentiment
with torch.no_grad():
    logits = model(**inputs).logits
    # Apply Softmax for 3-class probability distribution
    probs = torch.softmax(logits, dim=-1).squeeze().tolist()
    predicted_class_id = torch.argmax(logits, dim=-1).item()

SENTIMENTS = ["Negative", "Neutral", "Positive"]

print(f"Target Aspect : {target_aspect}")
print(f"Predicted Sentiment : {SENTIMENTS[predicted_class_id]}")
print(f"Confidence Scores   : Negative={probs[0]*100:.1f}%, Neutral={probs[1]*100:.1f}%, Positive={probs[2]*100:.1f}%")