Relation Extraction (RE) is one of the key technologies in the construction knowledge graphs. Currently, relation extraction methods based on BERT and neural networks typically rely on large-scale labeled data for pretraining. However, obtaining such large-scale labeled data in practical applications often comes with high costs. Although existing Large Language Models-based (LLMs-based) RE methods do not rely on labeled data. However, there are two main issues in LLMs-based RE methods: incomplete extraction of triples and incorrect triples extraction. To address these two issues in LLMs-based RE methods, we propose a new method: MCE. This method reduces the problem of incomplete triples extraction by using a 2-step extraction structure. Additionally, MCE employs a relation filtering module, a multi-choice extraction template, and an entity filtering module to eliminate erroneous triples. Experimental results show that, on the DUIE2.0 dataset, compared to the baseline GPT model (10-shot), MCE (1-shot) improved Recall by 15.02%, Precision by 7.84%, and F1-score by 10.3%. Notably, MCE even surpassed some full-shot models in Recall on the DUIE2.0 dataset, indicating that MCE is already capable of relatively complete extraction of triples on certain datasets.