Jigsaw Unintended Bias in Toxicity Classification

This is a kaggle competetion hosted by Jigsaw and google to identify toxic comments in online conversations. This is a addition to the previous Toxic Comment Classification Challenge which is to be more unbiased and diverse.

Methodology:

1. Data Cleaning:

   • For the training “comment_text” and “target” features were relevant. So, other columns are dropped.

   • For the training “comment_text” and “target” features were relevant. So, other columns are dropped.

2. Tokenizing and Embedding:

   • The training and testing data is tokenized and padded to have same length.
   • FastText and QuoraText embedding are combined, and fitted on tokenized dataset to create the embedding of the words.

3. Splitting the dataset:

   • The tokenized input variable and target variable is splitted for training and testing.

4. Model Building:

   • The model is built using Embedding layer, Bidirectional LSTM layer and Dense layer.

5. Model training:

   • The model is trained for 20 epoch with batch size of 248.
   • Early stopping and Reduce Learning rate is used to stop overfitting.
   • Model Checkpoint is used to save the best model.

Result Analysis:

• The training accuracy is: 0.9638 and loss is 0.0955
• The validation accuracy is: 0.96380 and loss is 0.095541
• The model seems converged well on the first epoch, the started overfitting. Reducing learning rate haven’t helped much.

Accuracy Plot:

Loss Plot:

Dependencies

• numpy
• pandas
• sklearn
• tensorflow
• matplotlib
• gensim
• tqdm
• operator
• gc

Acknowledgments

• Kaggle Competetion