A model and web app service in ABSA about smartphone review.
Data • Model • Web • Credits • Related • License
- We using the data collected by Luong Luc Phan et al., 2021, you can read more about the data when read their paper on this repo.
- Or visti the paper published on SpringerLink.
- We using PhoBERT-base model, which was created by VinAI for our problem and were inspired of fine-tuning by Quan Dang.
- You can read more details in this paper.
- With our model, we fine tune by use slow learning rate (1e-4) and small weight decay (5e-3) of AdamW. In the 3 last pooling layers, it performance look quite better than anything.
- You can visit our report for more information about the data, preprocessing and experiment setup.
- Best result:
| Tasks | Precision | Recall | F1-Score |
|---|---|---|---|
| Aspect Detection | 90.21 | 87.56 | 88.56 |
| Polarity Detection | 80.59 | 77.86 | 78.66 |
| Aspect + Polarity | 78.39 | 68.00 | 68.41 |
- 📝Predict random text from keyboard
- You can try our service by enter some review about phone before uploading file.
- By this, you can understand clearly our result.
- 🗂️Upload file
- The limit space of streamlit is 200MB. We sorry about that.
- From now, our model accepts 3 type of file: csv, json and excel.
- The processing for file input may take time depending on size of file.
To clone and run this application, you'll need Git and Pip installed on your computer. From your command line:
# Clone this repository
$ git clone https://github.com/leanhtu-AI/Sentiment-Analysis.git
# Go into the repository
$ cd Sentiment-Analysis
# Install dependencies
$ pip install requirements.txt
$ python train.py
Tip
I have setup for training models with gpu already, if you don't have gpu and still want training or fine tuning by yourself, you should go to this notebook or download weights here
$ streamlit run app.pyTip
When you running the first time, it will download the model weights(.h5) to your local filesystem. For faster running next time, you should create a folder named 'model' and move the h5 file to that. Then go to predict.py comment the downloaded model weights on cloud and uncomment the load weights local.
Note
We use the Windows command, if you are running on MacOS or Linux, you might to modify because it can be different a bit.
This repo uses the following open source packages:
- Tensorflow
- Streamlit
- Transformers
- Underthesea
- Others: nltk, Regex, Numpy,Pandas, Matplotlib, Seaborn,etc.
Junior-VB-web - Web version link of our model.
Important
We encourage you to run this by streamlit on local becaue the web is not always available on streamlit cloud.
MIT