Quantum-Inspired Cross-Attention Alignment for Turkish Scientific Abstractive Summarization

This paper presents a quantum-inspired cross-attention alignment approach for abstractive summarization. The motivation is that current neural summarizers often lose key content and produce summaries that are weakly grounded in the source, especially for long and information-dense scientific articles in low-resource languages. The method itself is model-agnostic and aims to strengthen token-level alignment without introducing additional trainable parameters or inference overhead, by exploiting a Born-rule-based similarity between encoder and decoder states. This general idea is instantiated and tested on the task of summarizing Turkish scientific articles in Mathematics Education, which provides a challenging low-resource test bed with long and dense source texts. Six different fine-tuning variants built upon the mBART-50 model are examined, including SFT, LoRA baselines, and two novel quantum-augmented decoders: the parameter-free SFT + QDA + QKernel and SFT + QDA + QBorn (Born-rule-inspired, learnable classical mapping). Models are trained with five random seeds and evaluated using beam search and sampling schemes. Statistical significance is assessed via bootstrap confidence intervals, Benjamini–Hochberg FDR correction, and Cliff’s δ effect size. Beam search consistently outperforms sampling across all architectures. Our best configuration, SFT + QDA + QKernel, achieves strong results (ROUGE-L: 0.2966, BERTScore-F1: 0.8890) and yields statistically significant, large-effect gains over all baselines. These findings indicate that the proposed parameter-free quantum kernel provides a practical way to improve abstraction quality and faithfulness, particularly in low-resource summarization settings.