Interventional Attribution of Semantic and Surface Novelty for Open-World Text Classification

This preprint presents an intervention-based attribution framework for open-world text classification that distinguishes surface-level novelty from semantic-level novelty. The approach uses controlled views of the same input to identify whether apparent novelty is caused primarily by changes in wording or by genuinely unseen semantic categories. The framework combines cause-aware routing, verification-aware candidate discovery, few-shot prototype learning, and compact continual updates. It is evaluated on BANKING77, CLINC150, HWU64, StackOverflow, and 20 Newsgroups, including tests of openness, backbone consistency, intervention fidelity, and transfer beyond intent classification. The results show strong surface-shift recovery, candidate ranking, and incremental learning on intent datasets, while also identifying limitations in calibration, domain transfer, and few-shot new-class learning.