Vietnamese Stock Sentiment Analysis (PhoBERT)
Model Description
This model is a Vietnamese financial sentiment analysis model fine-tuned from vinai/phobert-base for classifying the sentiment of Vietnamese financial news and stock-related text.
The model predicts three sentiment classes:
- NEG – Negative sentiment
- NEU – Neutral sentiment
- POS – Positive sentiment
It is designed to analyze Vietnamese financial news, stock market reports, and investment-related text.
Model Details
- Developed by: Minh Nguyen
- Model type: Sequence Classification
- Base model: vinai/phobert-base
- Language: Vietnamese
- Task: Sentiment Analysis
- Number of labels: 3
Intended Use
Direct Use
The model can be used to classify sentiment in:
- Vietnamese financial news
- Stock market analysis
- Investment reports
- Financial social media posts
Example applications:
- Market sentiment monitoring
- Financial news analysis
- Algorithmic trading signals
- Sentiment-based stock screening
Out-of-Scope Use
This model is not designed for:
- General Vietnamese sentiment analysis outside finance
- Non-Vietnamese languages
- Legal or medical text
- Critical decision-making systems without human oversight
How to Use
Install required libraries:
pip install transformers torch pyvi
#Example inference:
!git clone https://huggingface.co/mnguyn11/phobert-stock-sentiment-PTDLW
import sys
import os
# Add the directory containing 'predict.py' to the Python path
project_dir = "/content/phobert-stock-sentiment-PTDLW"
if project_dir not in sys.path:
sys.path.append(project_dir)
from predict import predict
predict(" Trong bối cảnh giá chung cư nội đô tăng mạnh và duy trì ở mức cao, dòng tiền đang dần chuyển sang các dự án còn nhiều dư địa tăng trưởng.")
Training Data
The model was fine-tuned using a dataset of Vietnamese financial news and stock-related sentences.
Dataset characteristics:
Domain: Finance / Stock Market
Language: Vietnamese
Labels:
NEG
NEU
POS
Approximate dataset size: ~1,200 labeled samples
Training Hyperparameters
Base model: vinai/phobert-base
Task: Sequence Classification
Number of labels: 3
Learning rate: 2e-5
Epochs: 3
Batch size: 16
Max sequence length: 256
Evaluation
Metrics
Evaluation metrics include:
Accuracy
Precision
Recall
F1-score
Results Model Accuracy Base PhoBERT ~33% Fine-tuned Model ~90%
The fine-tuned model significantly outperforms the base PhoBERT model on the financial sentiment classification task.
Limitations
The dataset size is relatively small (~1k samples)
The model is domain-specific to financial text
Performance may degrade on informal or slang-heavy Vietnamese
Environmental Impact
Training was conducted on a consumer GPU environment for experimentation purposes.
Acknowledgements
This model is based on the Vietnamese language model:
PhoBERT developed by VinAI Research.