A character level convolutional bilstm for arabic dialect identification

Abstract

In this paper, we describe the contribution of CU-RAISA team to the 2019 Madar shared task 2 1, which focused on Twitter User finegrained dialect identification. Among participating teams, our system ranked the 4th (with 61.54%) F1-Macro measure. Our system is trained using a character level convolutional bidirectional long-short-term memory (BiLSTM) network trained on approximately 2k users' data. We show that training on concatenated user tweets as input is further superior to training on user tweets separately and assign user's label on the mode of user's tweets' predictions.