Kaggle Quora Insincere Questions

Refer to the Kaggle competition description to get up to speed with the goal of this text classification competition. This competition is a kernel competition, meaning we can only use the training/test data and pre-trained embedding provided by the competition organizers.

Documentation

• pytorch_quora_insincere.ipynb [nbviewer][html] Four main personal learnings from this competition are:

  • Pre-trained embeddings helps with model performance if we leverage them correctly. Specifically, if we plan on using pre-trained embeddings, we should aim to get our vocabulary as close as possible to our pre-trained embedding.

  • Bucketing greatly improves the runtime. Meaning, instead of padding the entire input text data to a fixed length, we only pad the length for each batch.

  • How to use attention layer in the context of text classification.

  • Leverage pytorch framework for text classification.

Results

Private Leaderboard Score, F1 score:

• this repo: 0.69

• leaderboard score 850th place: 0.69

• leaderboard score 1st place: 0.71

Note that the model here is not extensively tuned and no ensembling/blending/stacking was used.