The goal of this ML analyser is to provide a system that can solve the following problems:
- Train your own sentiment analysis model.
- Provide an API endpoint to integrate the system with third-part services.
- Have a web-interface to interact with the analyser online.
The analyser is built using the Amazon Customer Reviews Dataset, in particular game reviews.
Multiple models have been built to server different needs, including the binary model and fine-grained ones.
The analyser supports at least 2 languages: Ukrainian and English but can be easily trained to work with other languages is it's built using statistical methods.
The overall performance of the system doesn't correspond to the production-ready expectations (mainly due to a lack of computational power during the model training) but may server as a foundation for the creation of more complex models.
To train a model, examine code in file ./examples/train_model.py
Here are a couple of confusion matrices for models built using linear regression.
-
Three classes (Positive, Neutral, Negative)
Accuracy: ~71%
-
Binary (Positive, Negative)
Acuracy: ~92%
