Data preprocessing is a core step in knowledge graph construction, consisting of two main stages: data collection and information extraction. This paper systematically reviews mainstream data preprocessing methods based on rules and lexicons, statistical machine learning, and deep learning, and thoroughly analyzes their technical principles and application limitations in entity recognition and relation extraction. Existing methods rely heavily on manual rules and suffer from weak semantic generalization, making it difficult to achieve cross-domain knowledge transfer. To address these issues, this paper explores a novel paradigm of “semantic-driven and automated extraction” based on large language models. By generating deep semantic embeddings through pre-trained large language models and combining vector similarity computation, it enables unsupervised and context-aware information extraction, driving the intelligent transformation of knowledge graph construction. The current approach is still in the exploratory stage, facing challenges such as high computational cost and low interpretability. Future research should focus on lightweight model design, multimodal semantic alignment, and domain knowledge integration to improve the efficiency of knowledge graph construction and model interpretability.