A Federated Approach to Predict Emojis in Hindi Tweets

Anonymous

Abstract: The use of emojis provide for adding a visual modality to textual communication. The task of predicting emojis however provides a challenge for computational approaches as emoji use tends to cluster into the frequently used and the rarely used emojis. Much of the research on emoji use has focused on high resource languages and conceptualised the task of predicting emojis around traditional servers-side machine learning approaches, which can introduce privacy concerns, as user data is transmitted to a central storage. We show that a privacy preserving approach, Federated Learning exhibits comparable performance to traditional servers-side transformer models. In this paper, we provide a benchmark dataset of $118$k tweets (augmented from $25$k unique tweets) for emoji prediction in Hindi and propose modification to the CausalFedGSD algorithm aiming to balance model performance and user privacy. We show that our approach obtains comparative scores with more complex centralised models while reducing the amount of data required to optimise the models and minimising risks to user privacy.

Paper Type: short