phobert-base_SA_finetuned

PhoBERT Base for Sentiment Analysis (Vietnamese)

1. Model Description

This model is a fine-tuned version of vinai/phobert-base for general Sentiment Analysis (SA) on Vietnamese text. It is designed to classify sentences into three emotional categories: Negative, Neutral, and Positive.

• Language: Vietnamese
• Model Type: Transformer-based Encoder
• Base Model: vinai/phobert-base
• Task: Sentiment Analysis (3-class classification)

2. Training Data

The model was trained on a comprehensive collection of Vietnamese sentiment datasets to ensure robustness:

• UIT-VSFC (ura-hcmut/UIT-VSFC)
• Vietnamese Sentiment Analysis (anotherpolarbear/vietnamese-sentiment-analysis)
• Kaggle Vietnamese Sentiment Analyst (linhlpv)

3. Training Procedure

Hyperparameters:

• Learning Rate: 5e-5
• Batch Size: 32
• Precision: Mixed precision (fp16)
• Epochs: 3
• Optimizer: AdamW

4. Evaluation Results

The model was evaluated on a dedicated test set (3166 samples).

• Accuracy: 93.90%
• F1-Macro (Average): 83.00%
• F1-Weighted: 93.59%

Classification Report:

Observations:

• The model performs exceptionally well on Negative (F1: 0.96) and Positive (F1: 0.95) sentiments.
• The Neutral class remains a challenge (F1: 0.59) due to the inherent ambiguity of neutral expressions in Vietnamese reviews and smaller support size.

5. How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("ZycckZ/phobert-base_SA_finetuned")
model = AutoModelForSequenceClassification.from_pretrained("ZycckZ/phobert-base_SA_finetuned")

text = "Sản phẩm dùng rất tốt, giao hàng nhanh chóng."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

predicted_class_id = logits.argmax().item()
# Map: 0: Negative, 1: Neutral, 2: Positive