Bridging Statistical and Contextual Representations: A Hybrid TF-IDF-MARBERT Framework for Arabic Emotion Classification

Abstract

Abstract

Emotion classification in Arabic text is constrained by morphological richness, dialectal diversity, and limited annotated resources. This study compares nine text representations across fifteen machine-learning classifiers on the SemEval-2018 Arabic emotion corpus (4,381 tweets, twelve classes). MARBERT with a stacking classifier gives the strongest single representation (accuracy 0.4413, F1 0.4115). A proposed hybrid representation, concatenating TF-IDF with MARBERT embeddings, raises accuracy to 0.6475 and F1 to 0.5889, a 20.6-percentage-point gain that compares favorably with recently reported large-language-model results on the same benchmark. Four classifiers that assume non-negative input fail on the hybrid representation's signed dimensions. The results indicate that lexical-contextual feature fusion is an effective, low-cost route to closing the performance gap left by any single Arabic text representation.