Abstract In recent successes of deep learning in the many fields of natural language processing, past studies of emotion recognition of Twitter user for the most part fo- cused around the utilization of lexicons and basic classifiers on pack-ofwords mod- els. The central question of this study is whether it can enhance their performance utilizing deep learning. To this end, it exploits hash tags to make three extensive emotion-labeled data sets comparing to various orders of emotions. At that point analyze the performance of a few word and character-based repetitive and convo- lutional neural systems with the performance on pack of-words and latent semantic indexing models. Moreover details check the transferability of the last hidden state representations between various classifications of emotions and whether it is con- ceivable to assemble a unison model for predicting every one of them utilizing a common representation. It is demonstrate that repetitive neural systems, particularly character-based ones, can enhance over pack of-words and latent semantic indexing models. While classify the tweet emotion, semantic of the token needs to be con- sidered. Semantic of the token stored in hash map to get searched easily. In spite of the fact that the exchange abilities of these models are poor, the recently proposed training heuristic delivers a unison model with execution similar to that of the three single models.Keywords: : Sentimental Analysis, Emotion Recognition, Twitter, Social Media, Hashtag, Tweet.