A multi-scenario approach to continuously learn and understand norm violations

Using norms to guide and coordinate interactions has gained tremendous attention in the multiagent community. However, new challenges arise as the interest moves towards dynamic socio-technical systems, where human and software agents interact, and interactions are required to adapt to changing human needs. For instance, different agents (human or software) might not have the same understanding of what it means to violate a norm (e.g., what characterizes hate speech), or their understanding of a norm might change over time (e.g., what constitutes an acceptable response time). The challenge is to address these issues by learning to detect norm violations from the limited interaction data and to explain the reasons for such violations. To do that, we propose a framework that combines Machine Learning (ML) models and incremental learning techniques. Our proposal is equipped to solve tasks in both tabular and text classification scenarios. Incremental learning is used to continuously update the base ML models as interactions unfold, ensemble learning is used to handle the imbalance class distribution of the interaction stream, Pre-trained Language Model (PLM) is used to learn from text sentences, and Integrated Gradients (IG) is the interpretability algorithm. We evaluate the proposed approach in the use case of Wikipedia article edits, where interactions revolve around editing articles, and the norm in question is prohibiting vandalism. Results show that the proposed framework can learn to detect norm violation in a setting with data imbalance and concept drift.