Geological Disaster Named Entity Recognition with Small Samples Based on Data Augmentation and Prompt Engineering

This paper uses a large language model to perform generative data enhancement on the original small sample data by performing random synonym replacement and random mask filling operations. In accordance with the reasoning logic of the large language model, three prompt templates are designed and the reasons are explored. Experiments show that when the parameters remain unchanged, the data enhanced by this method has been greatly improved under the three prompt templates, alleviating the difficulty of low resources of geological disaster data. And by comparing the performance of different instructions under different learning rates, the fine-tuning learning rate range suitable for the field of geological disasters is summarized. The limitation is that it is constrained by local computing resources, which reduces the parameter scale of LLM, and the recognition performance is low for extremely long or complex texts.