NBCG: Nash-Bargained Causal Game for Long-Tailed Multi-Label NLP

Long-tailed multi-label text classification is often treated as a data scarcity problem, addressed by re-sampling or fixed re-weighting. We argue that a central failure mode is \textit{dominant coalition capture}: frequent labels, amplified by spurious co-occurrences, form dominant coalitions that dominate shared representations and gradient allocation during optimization. As a result, rare labels are learned via superficial shortcuts, yielding brittle generalization under distribution shifts. We propose \textbf{NBCG}, a Nash-Bargained Causal Game that reformulates multi-label learning as a cooperative bargaining process among label coalitions. NBCG first leverages Neural Structural Equation Models to learn a directed dependency structure, inducing causally coherent coalitions---rather than random partitions---and coalition-specific communication masks. We then optimize a Nash bargaining objective over coalition utilities relative to an adaptive disagreement point, which serves as a principled credit-allocation mechanism: it adaptively prioritizes under-served coalitions while maintaining a Pareto-efficient trade-off among all players.

Many real-world text classification tasks require assigning several labels to the same document. However, AI models often learn common labels much better than rare ones. For example, if some categories appear very frequently in the training data, the model may focus too much on them and fail to recognize less common but still important categories. This can be harmful in applications such as medical coding, intent detection, or document organization, where rare labels may carry important information. This paper proposes a new training approach that helps different groups of labels receive fairer attention during learning. Instead of simply increasing the weight of rare labels by hand, our method automatically identifies related labels and adjusts training so that under-served groups are not overwhelmed by dominant ones. Experiments on several text classification benchmarks show that the proposed method improves performance on rare labels while keeping the model efficient and stable.