Food Safety News Summarization Paradigm Based On Large Language Models And RAG

The issue of food safety is becoming increasingly severe, and generating accurate and understandable news summaries is crucial for raising public awareness and supporting decision-making. However, current summarization methods lack effective integration of domain-specific knowledge and often suffer from issues such as factual errors, omission of key information, or misinterpretations, failing to meet the demand for high-quality news summaries in the food safety domain. This paper proposes a food safety news summarization generation paradigm based on the combination of Large Language Models (LLMs) and Retrieval Augmented Generation (RAG). First, the model extracts the "5W1H" (Who, What, When, Where, Why, How) from the news, providing key information for subsequent classification and knowledge retrieval. Then, based on the extracted "5W1H", the model classifies the article and narrows the knowledge retrieval scope, improving retrieval accuracy and ensuring the acquisition of domain-relevant knowledge. Finally, combining the original text, extracted "5W1H", and retrieved knowledge, the model generates the final summary, ensuring the summary accurately reflects the core content of the news and integrates domain knowledge. This paper constructs a knowledge base containing 4500 items of food safety knowledge and applies it to a self-built news data set from domestic authoritative food safety websites. Experimental results show that the proposed method has significant advantages in food safety news summary generation, especially in terms of relevance and information accuracy, performing better than traditional methods. Compared with the reference summary, the BLEU-4 is improved by 2.63% and ROUGE-L is improved by 1.67%. This approach is not only applicable to the food safety field but also has potential applications in automatic summary generation for other vertical domains.