Twitter sentiment analysis technology provides methods for polling public opinion on events or products. The majority of current research is aimed at obtaining sentiment features by analyzing lexical and syntactic features. Word embeddings express these characteristics explicitly. This paper proposes a new approach called Improved Words vectors for Twitter sentiments analysis (IWVTSA) to improve the f1 score of twitter sentiment analysis. We introduce a word embeddings method called tweet2vec, which was launched concurrently with doc2vec, to form a sentiment feature set of tweets; these word embeddings are combined with word sentiment polarity score features. For training and predicting sentiment classification labels, the feature set is integrated into four machine learning classifiers (XGBoost, SVM, Logistic Regression, Random Forest). We compare the quality of our model to that of the baseline model, the bag of words algorithm, and the results show that the combination of tweet2vec and Polarized lexicon in our model performs better on the F1-measure for Twitter sentiment classification.