Neural Network built with PyTorch for the Kaggle competition Quora Insincere Questions Classification. It finished in the top 2% (silver medal) with a F1-Score of 0.70445 (F1-Score of the winning model was 0.71323).
A Binary Classifier (toxic question or not), this model uses pre-trained embeddings (GloVe, Paragram and fastText) and blends with some statistical extra features the outputs from a bidirectional LSTM and a bidirectional GRU, using Attention layers, Pooling layers, and a Capsule layer.
Some of the machine learning techniques applied in the model are: Meta-Embeddings by Averaging, Cyclic Learning Rate and Threshold Optimization.
Kaggle Profile: https://www.kaggle.com/andrewparr