Insult detection using a partitional CNN-LSTM model

Abstract

Recently, deep learning has been coupled with notice-able advances in Natural Language Processing related research. In this work, we propose a general framework to detect verbal offense in social networks comments. We introduce a partitional CNN-LSTM architecture in order to automatically recognize verbal offense patterns in social network comments. Specifically, we use a partitional CNN along with a LSTM model to map the social network comments into two predefined classes. In particular, rather than considering a whole document/comments as input as performed using typical CNN, we partition the comments into parts in order to capture and weight the locally relevant information in each partition. The resulting local information is then sequentially exploited across partitions using LSTM for verbal offense detection. The combination of the partitional CNN and LSTM yields the integration of the local within comments information and the long distance correlation across comments. The proposed approach was assessed using real dataset, and the obtained results proved that our solution outperforms existing relevant solutions.