Multi-stage training and semantic fusion for cross-language summarization

Regarding cross-language automatic summarization research, conventional approaches have inherent problems such as semantic drift, alignment difficulties among multiple languages, different cultural loads, and non-parallel corpora complexities. To address these issues, this study proposes a cross-language summarization model that integrates multi-stage training and semantic fusion. The method uses a multilingual text-to-text transformer as the core architecture. It applies contrastive learning to reduce the semantic distance between the source language and the target language at the representation level, which improves cross-language alignment accuracy. On this basis, the model employs multi-head attention to combine context information from various levels and dynamically weight features for better feature representation and efficient computation. Besides, the approach applies knowledge distillation with a teacher-student architecture to obtain a lightweight model with fast inference speed. At the same time, it adopts a sentence-level semantic representation model for extracting semantic features to construct structured summaries, which makes the process of semantic fusion and generation control more precise. Experimental results prove that this proposed approach provides superior performance compared with current approaches in terms of generating better summaries on cross-lingual summarization datasets. The highest ROUGE-1 reaches 53.23%, and the highest ROUGE-L reaches 49.12%. The end-to-end generation latency is as low as 39.6 ms. The model also maintains competitive semantic consistency across the evaluated language directions and dataset settings. The experimental results provide evidence for the effectiveness of MSFS on multilingual cross-language summarization benchmarks, while its performance in low-resource languages, culturally loaded texts, unseen domains, and practical deployment requires further validation.