Abstract

Perceiving emotional state of human written text on tweeter also known as X depends upon textual characteristics by deep learning models but these models did not verify how many emojis are blended with text, what is the position of emoji and how many similar emojis are in row. In this study we test and verify that pattern of emoji in text and position either have any impact on model’s performance using SemEval-2018. This study employed logistic regression classifier in two ways, on text based features only as well as including emoji features blended with text. The Jaccard accuracy improved significantly to 0.2975 from 0.2875 on the dataset having emoji bearing tweets which is a improvement +0.0112 with  Whilst, the F1-micro score has increased from 0.4093 to 0.4722 relatively. The benefits of structural emoji encoding quantified by these results beyond slandered BoW approach.