Core Concepts
A privacy-preserving federated learning approach to train robust models for identifying offensive language online without compromising user privacy.
Abstract
This paper introduces a federated learning (FL) approach to address the challenge of training models for identifying offensive language online while preserving user privacy. The key insights are:
The authors propose a model fusion technique to combine multiple models trained on different datasets using FL, without the need to share the underlying data.
They evaluate the performance of the fused model on four publicly available English offensive language datasets (AHSD, OLID, HASOC, HateXplain) and show that it outperforms both non-fused baselines and ensemble models, while preserving privacy.
The fused model performs best when further fine-tuned on the same dataset used for evaluation, reflecting the ideal scenario where the model needs to perform well on the platform's specific data.
The fused model also generalizes well across different datasets, outperforming non-fused models evaluated on datasets other than the one used for training.
The authors also present initial multilingual experiments on English and Spanish, demonstrating the potential of the FL approach for low-resource languages.
Overall, the paper shows that federated learning is a promising approach for building privacy-preserving models for offensive language identification, outperforming traditional centralized training and ensemble methods.
Stats
"The spread of various forms of offensive speech online is an important concern in social media."
"Models trained to detect offensive language on social media are trained and/or fine-tuned using large amounts of data often stored in centralized servers."
"Federated Learning (FL) is a decentralized architecture that allows multiple models to be trained locally without the need for data sharing hence preserving users' privacy."
Quotes
"FL is a decentralized architecture that allows multiple models to be trained locally without the need for data sharing hence preserving users' privacy."
"We show that the proposed model fusion approach outperforms baselines in all the datasets while preserving privacy."