Few-shot Entity Extraction (Few-shot EE) is a task aiming to identify named entities via limited annotated samples. In recent years, the two-stage approach based on prototype networks has often been used to solve the Few-shot NER task, achieving remarkable results. However, false span and type confusion remain two shortcomings. To address the above shortcomings, an improved two-stage approach is proposed namely TSPN-KG: a Two-stage Prototypical Network for Few-shot Knowledge Graph Entity Extraction. It consists of span extractor stage and mention classifier stage. A method based on location calculation and span filtering is proposed to solve the false span problem in the span extractor stage, and another method based on information enhancement is proposed to solve the type confusion problem in the mention classifier stage. Experiments over multiple benchmarks demonstrate that the TSPN-KG outperforms strong baselines by a large margin.