Telecom VoC BanglaBERT — Sentiment Analysis

Fine-tuned BUET BanglaBERT (csebuetnlp/banglabert) for 3-class Sentiment Analysis (Positive, Neutral, Negative) across code-mixed Bengali, Romanized Banglish, and English telecommunications customer feedback.

Model Summary

• Base Architecture: csebuetnlp/banglabert (110M Parameters)
• Classes: Negative (0), Neutral (1), Positive (2)
• Validation Accuracy: 91.60% (Macro-F1: 0.8598)
• Corpus: 655,649 longitudinal customer reviews across Grameenphone, Robi, and Banglalink.

How to Use in Python

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="ratul-hasan/telecom-voc-banglabert-sentiment"
)

# Test review
print(classifier("নেট স্পিড একদম বাজে, সারাদিন বাফারিং করে"))
# Output: [{'label': 'Negative', 'score': 0.9821}]

Citation & Replication

GitHub Repository: telecom-Voice-of-Customer_intelligence