Refinement of Bengali Obscene Words using sequence to sequence RNNs

Abstract

In the recent past, a growing number of users have tolerated offensive behaviors or have witnessed objectionable activities through virtual platforms. Defamatory and abusive words are mostly used in the comment section on social networking sites that are more detrimental for children and teenagers. This type of offensive word can shape their thoughts in the wrong manners and influence their mental health. Our work targets to filter and replace the negative impression of defamatory words on children that are used on Facebook with the help of sequence to sequence RNNs along with LSTM. For doing this experiment, we have created a dataset by the comments and posts from Facebook. And also, done data pre-preprocessing, counting of vocabulary, counting of missing words, embedding words, finding the missing word, and so on. Our main focus was identifying the offensive word and filtering that word by the abstract text summarizer model and decreasing the training loss. After applying the model on the dataset, practically we have reached our target to decrease the training loss to 0.007 and are capable of making a filtered text from given input text.