28 April 2026 in IEEE Transactions on Multimedia
Existing Multimodal Aspect-Sentiment Pair Extraction (MASPE) methods primarily focus on aspect-sentiment pairs while only implicitly modeling aspect-object associations. However, they overlook explicit object-level prediction, which is crucial for accurate and interpretable sentiment analysis in realistic multimodal scenarios. To address this gap, we introduce the Multimodal Aspect-Object-Sentiment Triples Extraction (MAOSTE) task, which explicitly predicts aspect-object-sentiment triples while accommodating aspects without visual correspondence. We release the Twitter-MAOSTE and Twitter-MAOSTE-Political datasets, and propose the T5 Enhanced Aspect-Object-Sentiment Extractor (T5-AOSE), which augments the T5 framework by adding multimodal information and sentiment clues. Experiments show that T5-AOSE achieves strong performance on the challenging MAOSTE and MAOE tasks while maintaining competitive results on the traditional MASPE task, confirming its effectiveness and generalizability. The dataset and code are publicly available at Github