Cross-Language Multimodal Detection: A Novel Model for Identifying Duplicate Publications in Scientific Papers

This paper introduces a novel multimodal, cross-language detection model for identifying duplicate publications in academia, a problem where authors publish similar content across different languages without citation. Traditional detection methods are primarily text-based and struggle with linguistic variations and deliberately altered structures. Our model uniquely integrates textual, image, and structural modalities using a knowledge distillation SBERT for semantic similarity calculations between Chinese and English, and Vision Transformer (ViT) for image comparisons. We developed a new ChineseEnglish multimodal dataset in collaboration with the Chinese Journal of Electronics and Wanfang Data. On this dataset, our model has an AUC of 0.931 and an F1 of 0.82. The model's effectiveness is demonstrated through the identification and subsequent verification of two suspected duplicate publications. This research advances duplicate detection by combining multiple modalities and supports comprehensive academic content analysis.