Knowledge graphs, used for structured information representation, are growing rapidly with increasing data and users. This growth makes traditional machine learning models unsuitable for large graphs, highlighting the need for few-shot learning, which trains with minimal data. Few approaches evaluate false positives at random, regardless of entity resemblance. Sometimes, this leads to calculating an inaccurate similarity score without considering all available contexts. Existing approaches account for in-batch, pre-batch, and self-negatives, but they do not incorporate semantic data beforehand. Meanwhile, few other approaches emphasize the need for different hard negatives but do not employ translation scores. Our work addresses these specific aspects. This work introduces a novel methodology termed Similarity Aware Few-Shot Learning for Knowledge Graph Completion (SAFSL) to address these deficiencies. Initially, we propose a novel sampling method that aids in intuitively generating negative samples. We then propose an advanced scoring function that measures similarity among representations not only by considering cosine similarity but also by incorporating the translational property. The suggested model is further assessed utilizing the NELL dataset in comparison to existing methodologies. The findings indicate that the suggested model exhibits performance comparable to the existing state-of-the-art models.