This paper presents a system for automated subject tagging in a bilingual academic setting. Our approach leverages a novel burst attention mechanism to enhance the alignment between article and subject embeddings, derived from a large cross-lingual subject corpus. By employing a margin-based loss with negative sampling, our resource-efficient model achieves competitive performance in both quantitative and qualitative evaluations. Experimental results demonstrate average recall rates of 32.24% on the full test set, along with robust performance on specialized subsets, making our system well-suited for large-scale subject recommendation tasks.