Abstract
Federated Learning aims to learn machine learning models from multipledecentralized edge devices (e.g. mobiles) or servers without sacrificing localdata privacy. Recent Natural Language Processing techniques rely on deeplearning and large pre-trained language models. However, both big deep neuraland language models are trained with huge amounts of data which often lies onthe server side. Since text data is widely originated from end users, in thiswork, we look into recent NLP models and techniques which use federatedlearning as the learning framework. Our survey discusses major challenges infederated natural language processing, including the algorithm challenges,system challenges as well as the privacy issues. We also provide a criticalreview of the existing Federated NLP evaluation methods and tools. Finally, wehighlight the current research gaps and future directions.