Cryptonews sentiment classification model

This is a simple cryptonews sentiment classification Fast API.

Setting up the environment

To reproduce model training and running the API, perform the following steps:

• install Poetry with Python >= 3.8
• run poetry install to set up a virtual environment for the project and install all dependencies
• (opt.) poetry run pytest

Data

The training set is comprised of ~4500 news titles manually classiified into positive, neutral or negative. The data is not shared but you can refer to a toy dataset with 50 records: data/toy_train.csv.

title,sentiment
Bitcoin Market Has Run Out of Juice: Cryptocurrency Analyst,Negative
Bitcoin Core 0.14.0 Speeds Up Blockchain Syncing By 48%,Positive
Thinking of Travelling With Bitcoin? With These Websites You Can,Positive

Model

The model is a simple Tf-Idf & logreg Sklearn Pipeline. To reproduce model training, perform the following steps:

1. Check project configurations defined in config.yaml, e.g. model parameters and path to the model, etc.
2. Run poetry run python -m cryptonews_sentiment_base_model.train (might take a while for the first time, even with a small dataset)

API

Here we are using Fast API, first steps.

To run the API locally: poetry run uvicorn cryptonews_sentiment_base_model.api:app --reload

Then you can talk to the http://127.0.0.1:8000/classify endpoint sending text data as a dictionary with text field as a key and arbitrary text as a value, e.g.

{
   "title": "BTC drops by 10% this Friday"
}

Example of a Curl request:

curl --location --request POST 'http://127.0.0.1:8000/classify' \
--header 'Content-Type: application/json' \
--data-raw '{ "title": "BTC drops by 10% this Friday" }'

API docs

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc

Build docs

poetry run sphinx-build -a -b html docs/ docs/_build