Multi-dimensional Edge-Embedded GCNs for Arabic Text Classification

Ola Karajeh, Mohammed N. Al-Kabi, Edward A. Fox

Abstract: Text classification is considered an essential natural language processing task. However, Arabic text classification is challenging for many reasons, such as the complexity of the Arabic language, i.e., complex morphological structure, high degree of ambiguity that resulted from optional diacritics in the writing system, and multi-dialectical characteristics. Graph-based methods can capture global information by incorporating long-distance semantics and cross-document interdependencies. However, most graph neural networks are designed with single-value similarities. Edge features can provide rich information by using multi-dimensional edge embeddings. In this work, we introduce an Arabic Multidimensional Edge Graph Convolutional Network designed for text classification (AraMEGraph), that can capture more fine-grained linguistic details which are very important to provide a comprehensive understanding for such complex and ambiguous languages. Experimental results show that AraMEGraph significantly outperforms the state-of-the-art methods on three benchmark Arabic text datasets.

External IDs:doi:10.1007/978-3-031-72437-4_14