AASTE: Arabic Aspect Sentiment Triplet Extraction

Abstract

With the growing volume of Arabic user-generated content, there is a pressing need for sentiment analysis methods that address Arabic’s linguistic complexity. This paper presents Arabic Aspect-Sentiment Triplet Extraction (AASTE), the first span-based tagging model for extracting (aspect term, opinion term, sentiment) triplets from Arabic text. Built on AraBERT, the model employs novel 1D, 2D, and 3D tagging schemes to jointly identify sentiment-bearing components in restaurant reviews. It tackles challenges such as morphological richness and free word order using a tailored preprocessing pipeline and a greedy inference algorithm. Evaluated on a manually annotated dataset, the 3D scheme achieves 61.40% F1 overall, with strong performance in aspect (77.63%) and opinion (76.78%) term extraction, as well as single-token triplets (70.52%). The model offers a practical foundation for applications in business intelligence and customer feedback analysis. This work demonstrates the feasibility of adapting advanced sentiment analysis to morphologically complex languages like Arabic. <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>