Designing Semantic Computing for Social Good Applications with Long Tails

Many social systems, particularly those with high levels of interaction like online communities, often follow power law distributions, which include a significant ‘long tail.’ We argue in this position-paper on the criticality of researching and designing semantic computing architectures, especially those based on inductive Artificial Intelligence (AI) methods, that apply robustly to this long tail. The long tail is an important area of focus because it presents both technical and social challenges. Its inherent irregularity poses a test to the generalization abilities of current technology, including generative AI models like the large language models (LLMs). As machine learning and generative AI increasingly become part of user-facing systems, the complexities of the long tail must be addressed by both systems engineers and the AI community. We discuss two case studies with significant social impacts — detecting and fighting online human trafficking, and managing information in real-time during humanitarian crises — where AI and semantic computing designed for long-tail issues are crucially applied, often necessitating markedly different design approaches than those aimed at maximizing average accuracy.