kaggle_toxic_comments

This repository is a collection of models and scripts used for the 53rd Place Solution of the toxic comments challenge hosted on kaggle

TODOS:

• add models
• test models 1 by 1
• add preprocessing scripts

Model Explanations and references

GRU ATT

RNN with GRU cells and Average weighted attention layer

GRU CAPS

RNN with GRU cells topped with Capsules

https://arxiv.org/abs/1710.09829

HAN

Hierarchy attention network

https://www.cs.cmu.edu/~hovy/papers/16HLT-hierarchical-attention-networks.pdf

HLSTM

Hierarchy attention network with LSTM cells

NBSVM

Naive Bayes Logistic Regression

XGB

Gradient boosting with the xgboost package

https://xgboost.readthedocs.io/en/latest/

LGB

Gradient boosting with the light gradient boost package

Instructions

Folder structure

• assets
  • train.csv
  • test.csv
  • sample_submission
  • glove.twitter.27B.200d.txt (GloVe embeddings)
  • crawl-300d-2M.vec (fasttext embeddings)