Research on Data Construction Method of Entity Relation Instruction for Biological Invasion

Aiming at the key information extraction challenges of literature in the field of biological invasion, an entity-relationship data extraction method LRIE is proposed for low-resource scenarios. The method is centered around the semantic comprehension and generation capabilities of a large language model, and adopts a modularized pipeline architecture to process the tasks. First, predefined entity-relationship task labels are used, task examples are injected into the context based on the labels, and hint templates are constructed to enhance task interpretability. Second, a temperature-regulated multisampling strategy is introduced to filter unreliable results, mark fuzzy labels according to confidence level, and process fuzzy labels with multi-model voting. Finally, a structured parser is designed to convert model outputs into structured labeled data. Through the benchmark dataset comparison experiments, the LRIE method achieves a 1.1% improvement in F1 value, and the instruction data constructed by the LRIE method can better fine-tune the model so that the model obtains domain-adaptive characterization capability, and the comprehensive performance exceeds that of the existing baseline model.