Mizo Sentiment Classifier (hillbyte/mizo-sentiment-mizbert)

A state-of-the-art 3-class sentiment classifier for the Mizo language (lus / Lushai), fine-tuned on the pretrained robzchhangte/MizBERT (12-layer, 768-dimension, 110M parameter BERT architecture).

---

Model Details

• Model Type: BERT for Sequence Classification (num_labels=3)
• Base Architecture: robzchhangte/MizBERT
• Supported Language: Mizo (lus)
• Classes:
  • 0: Negative
  • 1: Neutral
  • 2: Positive

---

Evaluation Benchmarks

Evaluated on a stratified 15% hold-out test set (693 unseen samples) from combined Mizo sentiment datasets:

Metric	Score
Validation Accuracy	90.62%
Weighted F1-Score	0.9112
Macro F1-Score	0.8236

Per-Class Performance Breakdown

Class	Precision	Recall	F1-Score
Negative (0)	93.4%	90.7%	0.920
Neutral (1)	51.4%	73.5%	0.605
Positive (2)	96.3%	92.9%	0.945

---

Training Data & Methodology

The model was fine-tuned on 4,620 unified Mizo sentences synthesized from two community datasets:

1. hillbyte/mizo-sentiment (1,220 sentences, 3-class)
2. Blue7Bird/Mizo_sentiment_dataset (3,400 sentences, binary)

Optimization Highlights

• Stratified Split: 85% Train (3,927 samples) / 15% Validation (693 samples).
• Class-Weighted Loss: Applied inverse-frequency loss weighting (Negative: 0.83, Neutral: 4.71, Positive: 0.63) to effectively capture minority Neutral sentiments.
• Optimizer: AdamW with learning rate $3 \times 10^{-5}$ and linear warmup.

---

Quickstart & Usage

1. Using Hugging Face pipeline

from transformers import pipeline

classifier = pipeline("text-classification", model="hillbyte/mizo-sentiment-mizbert")

# Positive prediction
result = classifier("Aizawl khawpui ruahmanna thar hi a tha hle mai, kan lawm e!")
print(result)
# [{'label': 'Positive', 'score': 0.9916}]

# Negative prediction
result = classifier("Hetiang thil duhawm lo leh tha lo lutuk hi kan duh lo tawp mai.")
print(result)
# [{'label': 'Negative', 'score': 0.9938}]

2. Using PyTorch Directly

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "hillbyte/mizo-sentiment-mizbert"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "Aizawl khawpui ruahmanna thar hi a tha hle mai, kan lawm e!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)[0]

id2label = model.config.id2label
predicted_id = torch.argmax(probs).item()

print(f"Sentiment: {id2label[predicted_id]} ({probs[predicted_id].item()*100:.2f}%)")

---

Citation

If you use this model or the underlying datasets in your research or application, please cite the following publication:

@article{Lalrinmawii2025DetectingAC,
  title={Detecting Abusive Comments in Mizo: A Machine Learning Approach for a Low-Resource Language},
  author={R Lalrinmawii and Robert Lalramhluna and Gunavathi R.},
  journal={2025 IEEE International Conference on Advanced Visual and Signal-Based Systems (AVSS)},
  year={2025},
  pages={1-6},
  url={https://api.semanticscholar.org/CorpusID:281244917}
}