Graph convolution-based techniques for pragmatic Arabic figurative language classification.

Figurative language, including euphemism and metonymy, presents significant challenges in natural language processing (NLP) due to its abstract and context-dependent nature, particularly in morphologically rich and low-resource languages like Arabic. This paper introduces a graph-based embedding framework for figurative language classification that captures both syntactic dependencies and semantic relationships using heterogeneous graphs. We propose a configurable pipeline that converts text into structured graphs incorporating lexical, morphological, and syntactic cues, enabling deeper semantic reasoning. These graphs are processed using various graph neural network (GNN) architectures-such as GAT, HANConv, and MixHopConv-designed to model complex linguistic interactions. The approach is evaluated on two Arabic-language tasks: euphemism and metonymy detection. Our results demonstrate that attention-based and multi-hop GNNs outperform both traditional baselines and state-of-the-art transformer models (e.g., AraBERT, XLM-RoBERTa), particularly in metonymy detection where topological cues are more pronounced. HANConv and GAT achieve the highest F1-scores across tasks, while models like GraphConv and SAGEConv offer stability across configurations. We also introduce a validated Arabic lexical ontology for enriching semantic graphs. Our findings highlight the potential of graph-structured embeddings for nuanced linguistic tasks and suggest future directions including cross-lingual transfer, ontology expansion, and application to additional figurative categories.