Predicting multi-label emojis, emotions, and sentiments in code-mixed texts using an emojifying sentiments framework.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Online communication increasingly uses emojis and code-mixed languages (e.g., English-Hindi).
• Accurately matching emojis to the sentiment and emotion of code-mixed text is challenging.
• Existing datasets and methods may not fully address multi-emoji prediction in this context.

Purpose of the Study:

• To introduce a novel task of predicting multiple emojis for English-Hindi code-mixed sentences.
• To propose and release the SENTIMOJI dataset, an extension of SemEval 2020 Task 9 SentiMix.
• To develop and evaluate an end-to-end framework leveraging the relationship between emotion, sentiment, and emojis.

Main Methods:

• Developed a multi-task learning framework utilizing modified Transformer encoders (linear transformations, RMS-layer norm).
• Incorporated Gated Linear Units and Fully Connected layers for joint prediction of emojis, emotion, and sentiment.
• Created and utilized the SENTIMOJI dataset for training and evaluation.

Main Results:

• The proposed multi-task framework significantly outperformed single-task approaches on the SENTIMOJI dataset.
• Demonstrated a strong correlation between sentiment, emotion, and emoji usage in code-mixed text.
• Confirmed that sentiment and emotion identification aids in accurate multi-emoji prediction.

Conclusions:

• The study establishes the efficacy of a multi-task approach for emoji prediction in code-mixed language.
• Highlights the inherent link between linguistic sentiment/emotion and emoji selection.
• Contributes to advancing NLP tools for sentiment analysis and emotion recognition in multilingual online communication.