A sentiment analysis API

The application has been deployed as an API using heroku. It takes in a statement from the user and returns either of two values. Positive or negative.

Website link: https://sentiment-analysis-api-project.herokuapp.com/docs

How to setup

• Import the libraries from requirements.txt
  It can be done by writing the following comand in the terminal
  pip install -r requirements.txt
• Run the main.py file to get the model and the tokenizer, one can tune the hyper-parameters as one wishes.
• Open the app.py file and run the following command in the CLI python -m uvicorn app:app --reload after which the application will be hosted on the local system and a URL will be provided
• One can post the provided URL and observe the results on their own

Approach

• As this was an NLP task one had to first clean the text and then find a way to embed them
• NLTK and regrex were used to clean the text and end up with the processed text
• The in-built keras tokenizer was used for embedding the words, one-hot encoding, BoW, word2vec were considered but they did not provide the desirable results
• As for the model, a bidirectional lstm model has been built with a dropout layer and a final layer with sigmoid activation for the task of binary classification
• FastAPI was used to make the restful API
• Heroku was used for deployment

Results

• Accuracy -> 0.8939
• Precision -> 0.9007
• Recall -> 0.8874
• F1 score -> 0.8940

Notes

• tensorflow-cpu has been put in requirements.txt instead of the traditional tensorflow as tensorflow contains packages for both GPU and CPU which causes the slug size to increase by a factor of lot. While running of local system consider installing tensorflow by running the following command pip install tensorflow if any issues are encountered.
• The slug file of API is quite large, so if the API does not load in the first try. Kindly refresh the page.

Thank you, for giving your valuable time to review my submission.