Built a multi-label classification model that detects inappropriate texts and help us to further categorize them.
We have used a kaggle dataset which contains Wikipedia comments labeled by human raters for inappropriate behaviour. The dataset has 6 labels: ’toxic’, ’severe toxic’,’obscene’, ’threat’, ’insult’ and ’identity hate’. We have a total of 2,23,286 data points in our dataset and we have divided it into: 80% training, 10% validation and 10% testing
set. Thus the training set has 1,78,626 samples and the validation and testing set have 22,329 samples each.
As we have a natural language processing problem, we have first performed text cleaning which involves: converting to lowercase,removing special characters,removing numbers,removing stop words,replacing contractions with their full forms and lemmitization.
For factorizing the text data we have used 3 feature extraction techniques: Bag of word, TF-IDF, word2vec.
- Naive Bayes
- Logistic Regression
- SVM
- Random Forest
- Neural Networks
Out of these, tuned Neural Network with feature engineering, text cleaning and oversampling gave the best results for us with a score of 0.7 (Using 60% Recall + 40% Precision).
For further details please check the final report.