Toxic Comment Classification

Classify Wikipedia comments into 6 toxicity categories: toxic, severe_toxic, obscene, threat, insult, identity_hate.

NLP ROC-AUC TF-IDF + LogReg (6 label)

Feature Importance

text_tfidf 0.40 word_count 0.12 char_count 0.10 caps_ratio 0.10 excl_count 0.08 unique_ratio 0.08 avg_word_len 0.07

Key Insights

Pre-trained embeddings matter most. Architecture changes barely help. Examining last 25-50 chars separately captures end-of-comment toxicity. Pseudo-labeling helps.