manuelemacchia/hotel-sentiment-analysis

Implementation of a data science process for predicting the sentiment contained in hotel reviews by building a binary classification model

Ciao, I'm Manuele 👋

• Machine learning engineer based in Bari, Italy.
• Working at Connect Reply.
• MSc in Data Science and Engineering at Politecnico di Torino.
• Interested in natural language processing, computer vision, generative AI, MLOps, web development, and anything to do with data.

manuelemacchia.com

👨‍💻

hotel-sentiment-analysis's Issues

Create GitHub page

Once we close #4 we should create a GitHub page for this project.

Handle repeating characters from the beginning and the end of a token

The spell checker currently removes characters that appear consecutively more than twice (e.g., pproovaaaa becomes pproovaa).

In addition to that, it should remove characters that repeat more than once from the beginning and the end of a token (e.g., pproovaaaa becomes proova).

This should help the stemmer to correctly stem tokens.

Handling text emoticons

Right now, the classifier can handle unicode emojis such as 🥰 or 😡.
We shall improve the emoji handler with support for text emoticons, such as :) or :(.

Test set unavailability should be solved by train test split on the development set

The test set is unavailable. Therefore, we should use the development set for training, testing and validation.

To do this, we should split the development set into training and test sets (80-20?). To validate the model, we should perform k-fold cross-validation on the training set. Testing should be done on the test set.

Modularize the classifier

We could modularize the classifier code. This requires further exploration: what modules? use python packages? documentation?

Write more extensive documentation in markdown format

Delete the PDF report once the markdown documentation is finished.

Review length distribution graph

We could graph the length of the reviews in the data exploration phase. We could have two distributions, one for each class (negative and positive review length distribution). Is this useful?