Scientist attribute extraction based on instruction supervised fine-tuning and loss function redefinition

The challenge of obtaining high-quality labeled data for scientist attributes significantly impacts traditional deep learning models in the task of scientist attribute extraction. These models typically depend heavily on extensive labeled datasets to achieve optimal performance. However, due to the limited availability of such high-quality labeled data, these models often exhibit weaker generalization abilities when applied to new, unseen data. Utilizing large language models, which possess stronger text comprehension abilities, for scientist attribute extraction can effectively reduce dependence on labeled data. However, the issue of "hallucination" in large language models affects their performance in this task. To address this issue, this paper firstly constructs an ontology for scientist attributes and designs prompt strategies. Subsequently, the idea of bounding box regression loss from object detection tasks is adopted to redefine the loss function for Fine-Tuning large models. This redefinition ensures a clear convergence direction when applying large models to scientist attribute extraction task, effectively reducing LLM hallucination and enhancing the accuracy of the attribute identification process. Our empirical evaluation demonstrates that this novel methodology substantially enhances the efficacy of large models in scientist attribute extraction task.