LLM fine-tuning for international news style generation and cross-cultural dissemination quality assessment

The rapid globalization of news media has heightened the demand for high-quality, culturally adapted international news content. However, generating news articles that faithfully conform to the stylistic conventions of diverse national media outlets while maintaining cross-cultural semantic fidelity remains a formidable challenge. This paper presents a novel framework that leverages fine-tuned large language models (LLMs) for international news style generation and introduces a comprehensive cross-cultural dissemination quality assessment (CDQA) methodology. Specifically, we construct a multilingual international news corpus spanning seven major languages and ten distinct journalistic style profiles, drawn from leading international media organizations. We then apply parameter-efficient fine-tuning via Low-Rank Adaptation (LoRA) on a LLaMA-2-based backbone, augmented with a style-aware prefix adapter that conditions generation on target-style embeddings. For quality assessment, we propose a multi-dimensional evaluation framework integrating automatic metrics—including BLEU-4, ROUGE-L, and BERTScore—with a novel Cross-Cultural Appropriateness Score (CCAS) grounded in culturological theory. Experimental results demonstrate that our fine-tuned model achieves a BLEU-4 score of 42.7 and a BERTScore F1 of 0.876 on held-out test sets, outperforming strong baselines by 5.3 and 3.1 percentage points, respectively. Human evaluation conducted by bilingual journalists across five language pairs yields mean fluency and cultural appropriateness ratings of 4.21/5.00 and 4.08/5.00, respectively. Cross-cultural quality classification accuracy reaches 87.3%, validating the effectiveness of the CCAS framework. This work offers a scalable, reproducible pipeline for high-fidelity international news generation, with direct implications for global media localization.