An Algorithm for Automatic Generation of Tags for Children’s Health Science Resources Based on Instruction-Tuned LLM and Similarity Retrieval

In this paper, the research goal is to realize the automatic generation of children’s health science resources, in order to solve the two problems of inaccurate generation of resource labels in the field and the difficulties in standardization of label generation using Large Language Model (LLM). This paper proposes an automatic label generation algorithm for children s health science resources based on instructed LLM and vector (or word embedding) similarity retrieval. The algorithm first performs command fine-tuning of pre-trained LLM based on children s health science resources and LoRA (Low-Rank Adaptation) technology. Chroma vector database is then used to vectorize the standardized labels. Finally, the label results of instructed LLM are similarly matched with the vectorized standardized labels to generate the standardized label results. This paper verifies the effectiveness of the algorithm through three sets of comparative experiments based on the original data of 10,000 children’s health science resources. The experimental results show that the automatic generation ability of children s health science resources labels has been improved after domain data fine-tuning LLM and similarity retrieval.