Arabic Multimodal Dataset for Aspect-Based Sentiment Analysis

Abstract

Multimodal sentiment analysis (MSA) has attracted growing interest among researchers in recent years, as multimodal user-generated content (e.g., text and images) has become increasingly common on the Internet. In Aspect-Based Sentiment Analysis (ABSA), multimodal data can provide complementary information to text by capturing aspect-related information across text and image modalities. However, to the best of our knowledge, no existing study has addressed Multimodal Aspect-Based Sentiment Analysis (MABSA) for Arabic user-generated content using paired text and images. To address this gap, we constructed and annotated Arabic Multimodal Aspect-Based Sentiment Analysis (AraMABSA), a new Arabic MABSA dataset comprising 1721 Arabic hotel reviews associated with 3653 images collected from Booking.com for hotels located in Saudi Arabia and covering six aspect categories. We developed a human annotation pipeline with well-defined annotation guidelines, structured into three stages adopted from prior MABSA research: text annotation, image annotation, and text–image annotation. We further conducted a systematic validation procedure to evaluate the quality and reliability of the annotated labels, using Inter-Annotator Agreement (IAA). Finally, we performed technical validation of the constructed dataset using Qwen3-VL-4B-Instruct with supervised Low-Rank Adaptation (LoRA) fine-tuning for Aspect Category Detection (ACD) and Aspect-Based Sentiment Polarity Classification (ASPC) under text-only and text+images conditions. The IAA results demonstrated strong agreement among the annotators, supporting the consistency of the annotation process and guidelines. The experimental results further showed that the dataset supports learnable ACD and ASPC tasks under the evaluated setup. Text+images increased Macro-F1 from 0.6865 to 0.8761 for ACD and from 0.8257 to 0.8837 for ASPC. For ACD, both input conditions were evaluated against the final text–image reference annotations, with the larger text+images advantage occurring for aspects added during text–image annotation. For ASPC, sentiment labels were derived from the text; therefore, the observed text–image improvement should be interpreted within this annotation design. These findings provide initial technical validation of the AraMABSA dataset and support its use as a resource for future Arabic MABSA research.