Topic Based Sentiment Detection using BERT

Jump to this document to understand the data, approach and further improvements

Stack

• transformers
• torch
• streamlit
• fastapi
• pandas

Setup

• Clone the repo
• Install requirements: conda create --name topicsentiment --file requirements.txt or pip install -r requirements.txt in your env
• Download pre-trained model files here
• Copy the files to model_dir in config

Running the demo

• Run uvicorn prediction_api:app
• Run streamlit run st_app.py
• Go to http://localhost:8501/

Training

Pandas and Torch dataset classes in topicsentiment/dataset.py

Model Class and Trainer Class in topicsentiment/model.py

Jump to topicsentiment for detailed info

Configuration in config.py

Run python train.py --data sentisum-evaluation-dataset.csv for full pipeline - preprocess and train

Run python train.py --data sentisum-evaluation-dataset.csv --preprocess saves the preprocessed dataset to model_dir in config

Run python train.py --data final_data.pkl --train trains the preprocessed dataset

Prediction API

Built using FastAPI

Code in prediction_api.py

Run uvicorn prediction_api:app

Swagger Docs at http://127.0.0.1:8000/docs

Streamlit Interface

Simple Interface built using streamlit

Code in st_app.py

Run streamlit run st_app.py

App at http://localhost:8501/

Planned Additions

• Alternative approaches
• Notebook for data exploration
• Docker image