Presentation Information
[2Yin-B-05]On the Role of Multimodal Graph Sparsification in Machine Translation
〇ZHENGPAN FEI1, YUTING ZHAO1 (1. Kyushu University)
Keywords:
MT
Graph-based structures are increasingly used to capture fine-grained interactions between textual and visual representations for multiple NLP tasks. However, these graphs are typically densely connected, introducing redundant cross-modal relations, increased computational cost, and potential noise propagation.
In this work, we investigate the role of graph sparsification in the multimodal machine translation task by exploring pruning strategies applied to multimodal graphs. Specifically, we vary pruning rates over cross-modal nodes and edges to analyze how different sparsity levels affect translation quality and model efficiency. Experimental results on the Multi30k benchmark in machine translation tasks on EN-DE and EN-FR demonstrate that moderate sparsification can preserve translation performance while reducing complexity cost.
In this work, we investigate the role of graph sparsification in the multimodal machine translation task by exploring pruning strategies applied to multimodal graphs. Specifically, we vary pruning rates over cross-modal nodes and edges to analyze how different sparsity levels affect translation quality and model efficiency. Experimental results on the Multi30k benchmark in machine translation tasks on EN-DE and EN-FR demonstrate that moderate sparsification can preserve translation performance while reducing complexity cost.