Toxic comment classification using convolutional and recurrent neural networks

This thesis aims to provide a reasonable solution for categorizing automatically sentences into types of toxicity using different types of neural networks. There are six types of categories: Toxic, severe toxic, obscene, threat, insult and identity hate. Three different implementations have been studied to accomplish the objective: LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit) and convolutional neural networks. The thesis is not thought to aim on improving the performance of every individual model but on the comparison between them in terms of natural language processing adequacy. In addition, one differential aspect about this project is the research of LSTM neurons activations and thus the relationship of the words with the final sentence classificatory decision. In conclusion, the three models performed almost equally and the extraction of LSTM activations provided a very accurate and visual understanding of the decisions taken by the network.