[EDIT:
-
Adding a more simple (and more efficient) third alternative method: NBSVM using a Kaggle kernel from fastai's Jeremy Howard.
-
Adding a BERT method, using the ktrain package. Got the best results, but keep in mind its computational cost (ran in Colab). There's still margin for improvement, it served mostly for experimenting purposes.]
Given a (non public) dataset with real social media comments from the page of an online store, the goal was to create a model to detect harmful comments, while applying some NLP techniques (using nltk and Gensim packages; emoji too 😛💥).
The data had the comments' text, detected language of the comment, its translation to English, the indication of the comment being harmful or not and the type of harmful comment.
Due to ⏳ issues, I decided to approach the problem in the most practical way I could, like choosing to work on the English translation of the comments only or limiting it to the harmful/not-harmful binary label.
It's an old exercise, but eventually I'll pick it up again to improve results and try different approaches. For now, two different ones: Random Forest and a Bidirectional LSTM approach⬅️➡️.