Incremental Machine Learning for Text Classification in Comment Moderation Systems

Wolters, Anna; Müller, Kilian; Riehle, Dennis Maximilian

Abstract

Over the last decade, researchers presented (semi-)automated comment moderation systems (CMS) based on machine learning (ML) and natural language processing (NLP) techniques to support the identification of hateful and offensive comments in online discussion forums. A common challenge in providing and operating comment moderation systems is the dynamic nature of language. As language evolves over time, continuous performance evaluations and resource-inefficient model retraining are applied to ensure high-quality identification of hate speech in the long-term use of comment moderation systems. To study the potentials of adaptable machine learning models embedded in comment moderation systems, we present an incremental machine learning approach for semi-automated comment moderation systems. This study shows a comparison of incrementally-trained ML models and batch-trained ML models used in comment moderation systems.