This paper proposes a model for the refining of word vectors that does not require a labelled corpus and may be applied to any pre-trained word vectors. The suggested technique employs a sentiment intensity lexicon that may offer realvalued sentiment data to rank a collection of semantically and emotionally related nearest neighbors for each word. The ranking nearest neighbors is then utilized to control the direction and distance of the refinement technique, which iteratively enhances the word vector representation of each word. Experiments using SST demonstrate that the suggested strategy outperformed conventional word embeddings and sentiment embeddings for both coarse-grained and binary sentiment categorization. In addition, the performance of several models of deep neural networks has been enhanced. In future study, the suggested approach will be evaluated on more datasets. Additionally, additional experiments will be undertaken to offer a more comprehensive study.