Vietnamese Student Feedback Sentiment Analysis

Project Overview

Universities receive tens of thousands of student reviews per semester. Manual analysis is impractical at scale. This project delivers an End-to-End NLP pipeline that automatically classifies Vietnamese student feedback into three sentiment polarities: Positive, Neutral, and Negative.

Why This Problem Is Hard

Vietnamese NLP presents unique challenges:

• Compound words — "giảng_viên" (lecturer) must be treated as one token, not two
• Tone-dependent semantics — Vietnamese uses diacritical marks that fundamentally change word meaning
• School-specific vocabulary — domain-specific terminology rarely covered in general models

Solution: Fine-tuning PhoBERT

PhoBERT is a state-of-the-art BERT-based model pre-trained specifically on a large Vietnamese corpus by VinAI Research. Fine-tuning it on domain-specific student feedback data yields exceptional understanding of educational Vietnamese.

Methodology

1. Data Preprocessing

from pyvi import ViTokenizer

def preprocess_text(text: str) -> str:
    # Word segmentation — critical for Vietnamese
    # e.g., "giảng viên" → "giảng_viên"
    return ViTokenizer.tokenize(text)

• Applied ViTokenizer for Vietnamese word segmentation
• Handled massive dataset of student feedback with label mapping
• Dynamic padding + truncation strategy (max length: 128 tokens)

2. Model Architecture

PhoBERT-base (vinai/phobert-base)
    ↓
[CLS] Token Representation
    ↓
Classification Head (Linear: 768 → 3)
    ↓
Softmax → {Positive, Neutral, Negative}

• Base: vinai/phobert-base via AutoModelForSequenceClassification
• Training: AdamW optimizer with learning rate warmup
• Hardware: NVIDIA L4 GPU (Google Colab)

3. Evaluation Results

Overall Accuracy: 94.6% on 3,100+ unseen test samples.

Key Design Decision: Prioritizing Negative Detection

The F1-Score of 0.96 for Negative feedback is the most critical metric: missing a genuine complaint costs universities actionable insight. This was achieved by carefully balancing the dataset and tuning the classification threshold.

Production Readiness

The inference pipeline is modular and API-ready:

def predict_sentiment(text: str) -> dict:
    # 1. Vietnamese word segmentation
    segmented = ViTokenizer.tokenize(text)
    # 2. Tokenize for PhoBERT
    inputs = tokenizer(segmented, return_tensors="pt",
                      max_length=128, truncation=True, padding=True)
    # 3. Inference
    with torch.no_grad():
        logits = model(**inputs).logits
    # 4. Return human-readable result
    label = ["Negative", "Neutral", "Positive"][logits.argmax()]
    confidence = torch.softmax(logits, dim=1).max().item()
    return {"label": label, "confidence": f"{confidence:.1%}"}

Wrappable directly into a FastAPI endpoint for a university analytics dashboard.

Technology Stack

• Deep Learning: PyTorch, HuggingFace transformers, datasets
• Pre-trained Model: vinai/phobert-base
• Vietnamese NLP: pyvi (ViTokenizer)
• Evaluation: Scikit-learn (classification report, confusion matrix)
• Data: Pandas, NumPy
• Hardware: NVIDIA L4 GPU (Google Colab)