A Method for Constructing Railway Transportation Knowledge Graphs Based on DeepSeek Model Fine-Tuning

To overcome the limitations of traditional methods for railway knowledge graph construction, such as high manual reliance, poor scalability, and ineffective unstructured text processing, this paper proposes an automated triplet extraction approach based on a supervised fine-tuned DeepSeek large language model. Using domain-specific prompts and LoRA, we fine-tune the 8B model with limited data, conducting comparative analyses across a range of training sample sizes and hyperparameters. While results confirm that the information extraction performance of general-purpose models exist a positive correlation with parameter size, our findings show that with limited samples, avoiding hyperparameters that cause drastic parametric changes is critical. And after domain-specific few-shot fine-tuning, the 8B model attained a triplet extraction F1-score of 0.8504, markedly outperforming the general-purpose online full-parameter model. Applied to railway transportation tasks which include information extraction, knowledge graph construction, and visualization, the optimally fine-tuned 8B model demonstrates the feasibility of efficient, accurate domain knowledge graph construction under low-resource conditions, offering effective support for knowledge management and intelligent applications in railway transportation.