Noise-Regularized Bidirectional Gated Recurrent Unit With Self-Attention Layer for Text and Emoticon Classification

Abstract

The emoji are capable of expressing emotion beyond the meaning of the text by displaying visual emotions, which makes the content more distinct. Recently, emoji and text prediction has gained more significance, since it is hard to choose the appropriate one from thousands of emoji candidates. The small-sized dataset provides a poor description of features that resulted in classification and showed overfitting and underfitting problems. Therefore, Noise Regularized Bidirectional Gated Recurrent Unit (Bi-GRU) with Self-Attention Layer (SAL) is proposed for the classification of text and emoji. The proposed Noise Regularized Bi-GRU which is an aspect-based sentiment analysis performs a series of experiments on Twitter data to predict the sentiment of a tweet. The proposed Noise Regularized BGRU with SAL method obtained an accuracy of 87.77 % better when compared to the deep learning model that obtained an accuracy of 86.27 %.