Natural Language Processing using Federated Learning: A Structured Literature Review
Gerrit Schumann, Jan-Philipp Awick, Jorge Marx Gómez · 2023
Federated learning (FL) addresses privacy concerns and data distribution challenges in machine learning, as it enables decentralized training on local devices without the need to share raw data. This can be particularly relevant for cross-company collaboration, where data privacy is paramount and shared learning can bring mutual benefits. Since text data is highly sensitive in many real-world use cases, there is great benefit in combining FL with natural language processing (NLP). Given the increasing amount of locally generated text data and the tightening of privacy requirements, this paper aims to analyze previous research on the use of federated learning for NLP. To this end, we conducted a structured literature review based on 76 publications to determine the NLP tasks that were addressed using FL, the types of text data and machine learning approaches used, and the setting of the FL-architectures. In this way, we have shown the current state of the art, identified potential challenges, and highlighted possible research directions.