English diversified and high-fidelity translation based on adaptive label smoothing algorithm

This study proposes a novel English translation method using an adaptive label smoothing algorithm to address issues of rigid and non-diverse outputs.It employs three modules: a grammar feature capture module (redefined as a 'multi-modal syntactic structure modeller'), a component attention-enhanced encoder, and an adaptive label smoothing decoder.These create an interaction mechanism of syntactic feature, attention, and loss weight flows to better represent sentence structures and generate varied translations.Experimental results show the model's repetition rate reduced to 0.09, beating diverse beam search (0.12).It achieved a syntactic tree similarity score of 0.87, outperforming a grammar-enhanced model by 0.09.With ten candidate outputs, it showed higher diversity in sentence length and n-gram differences (8.60 and 0.92) compared to baseline methods.The model also demonstrated efficient convergence, with loss dropping to 0.98 after 50 epochs.The method enhances translation accuracy, semantic fidelity, and output diversity under a clear evaluation framework.