Multi-Topic Labelling Classification Based on LSTM

Abstract

The necessity for automatic classification of some resources has become extremely important given the fast-increasing number of electronic resources. People's opinions have been extracted from social media sites using Artificial Intelligence (AI). Despite this, the majority of current research focuses on extrapolating features from texts. Multi-label textual data classification is a significant problem in terms of the increasing amount of data available and the growing difficulties of assigning each text piece with one label. Examples include news and email articles. This work focuses on multi-label classification of Arabic texts. After dataset collection; several architectures were tested for this task. Bidirectional Long Short-Term Memory networks (BiLSTM) showed the superior results with F-score equal 86.6 in development set, and F1-score equal 82.24 in leaderboard Mowjaz competition.