Authors - Duy Pham, Tung-Duong Le-Duc, Anh-Tai Pham-Nguyen, Trung Nguyen Mai, Long Nguyen, Dien Dinh Abstract - Multimodal knowledge graphs improve structured knowledge representation and tasks such as cross-graph entity alignment. However, most benchmarks focus on resource-rich languages and assume dense relational structures and balanced attributes. Low-resource languages like Vietnamese pose additional challenges, including structural sparsity, attribute asymmetry, and modality noise. To address this gap, we in troduce DBWiki-VN15K, the first large-scale Vietnamese multimodal knowledge graph dataset for entity alignment. Built from Wikidata and DBpedia, it contains 15,000 aligned entity pairs with relational triples, lo calized numerical attributes, and visual modalities. The dataset provides both word-segmented and unsegmented text to support different linguis tic processing approaches. Experiments with state-of-the-art multimodal entity alignment models reveal that structure-guided multimodal fusion and dynamic modality weighting are more robust to sparse and noisy features. Additionally, unsegmented subword tokenization better han dles cross-graph translation inconsistencies than strict Vietnamese word segmentation. DBWiki-VN15K offers a realistic benchmark for studying multilingual and multimodal knowledge fusion. Our dataset is available at: https://github.com/Tim50c/DBWiki-VN15K.
Open Zoom