Diagnosing Representation Dynamics in NER Model Extension (arxiv.org)

Researchers tested how to extend a BERT-based NER model trained on semantic classes (PER, LOC, ORG) to also recognize pattern-based PII (EMAIL, PHONE) in noisy spoken-language data. Joint fine-tuning showed surprisingly little degradation on the original semantic classes, suggesting the model can learn pattern-like PII without overwriting semantic representations. To probe why, the team used an incremental learning diagnostic to measure representation drift and uncovered two mechanistic insights. First, LOC labels are uniquely fragile because some location mentions share pattern-like signals (e.g., postal codes), creating representation overlap with PII and causing semantic drift. Second, they discovered a “reverse O-tag representation drift”: the model initially encodes PII patterns into the background “O” class, which blocks new PII learning. Only by unfreezing the O-tag classifier does the background class adapt and “release” those patterns, enabling correct PII acquisition. The study highlights a practical checklist for extending NER systems: treat semantic vs. morphological features separately, monitor class-wise representation overlap, and consider targeted unfreezing of the O-class or classifier heads. These findings offer actionable guidance for transfer/continual learning in NER, especially when adding pattern-driven entities to models deployed on noisy conversational data.