Detection of Toxic Content on Social Networking Platforms Using Fine Tuned ULMFiT Model

Abstract

Question and answer websites such as Quora, Stack Overflow, Yahoo Answers and Answer Bag are used by professionals. Multiple users post questions on these websites to get the answers from domain specific professionals. These websites are multilingual meaning they are available in many different languages. Current problem for these types of websites is to handle meaningless and irrelevant content. In this paper we have worked on the Quora insincere questions (questions which are based on false assumptions or questions which are trying to make a statement rather than seeking for helpful answers) dataset in order to identify user insincere questions, so that Quora can eliminate those questions from their platform and ultimately improve the communication among users over the platform. Previously, a research was carried out with recurrent neural network and pretrained glove word embeddings, that achieved the F1 score of 0.69. The proposed study has used a pre-trained ULMFiT model. This model has outperformed the previous model with an F1 score of 0.91, which is much higher than the previous studies.