A self-attention hybrid emoji prediction model for code-mixed language: (Hinglish)

Emojis are an essential tool for communication, and various resource-rich languages such as English use emoji prediction systems. However, there is limited research on emoji prediction for resource-poor and code-mixed languages such as Hinglish (Hindi + English), the fourth most used code-mixed language globally. This paper proposes a novel Hinglish Emoji Prediction (HEP) dataset created using Twitter as a corpus and a hybrid emoji prediction model BiLSTM attention random forest (BARF) for code-mixed Hinglish language. The proposed BARF model combines deep learning features with machine learning classification. It begins with BiLSTM to capture the context and then proceeds to self-attention to extract significant texts. Finally, it uses random forest to categorize the features to predict an emoji. The self-attention mechanism aids learning since Hinglish, a code-mixed language, lacks proper grammatical rules. The combination of deep learning and machine learning algorithms and attention is novel to emoji prediction in the code-mixed language(Hinglish). Results on the HEP dataset indicate that the BARF model outperformed previous multilingual and baseline emoji prediction models. It achieved an accuracy of 61.14%, precision of 0.66, recall of 0.59, and F1 score of 0.59.