Zefan Cai

Zefan Cai

PhD student at UW-Madison. Test-time training, efficient architecture, long-context memory.

SANTA: Separate Strategies for Inaccurate and Incomplete Annotation Noise in Distantly-Supervised Named Entity Recognition

← All research

Language Understanding

Findings of ACL 2023 · PDF · Code · Data

Give wrong labels and missing labels different treatments. Distant supervision can assign an entity the wrong type or miss it entirely. SANTA connects inaccurate annotations to entity ambiguity, and incomplete annotations to a shifted decision boundary, then designs a separate strategy for each problem.

SANTA architecture with separate training strategies for incomplete and inaccurate annotations and entity-aware KNN at inference.
Figure 2: SANTA treats missing entities with boundary mixup and a robust loss, and mislabeled entities with memory smoothing and entity-aware retrieval. Click to enlarge.

For spans labeled as entities, memory-smoothed focal loss stabilizes learning and entity-aware KNN supports disambiguation at inference. For spans labeled as non-entities, boundary mixup and a noise-tolerant loss help recover entities missed by the knowledge base. The two branches share a span-based NER encoder.

The experiments cover five distantly supervised NER datasets: CoNLL2003, OntoNotes5.0, Webpage, BC5CDR and EC. SANTA improves over the compared baselines across these datasets; on Webpage it reports 71.79 F1, versus 62.98 for a baseline that separates the losses but omits the specialized modules. The method and noise analysis are specific to this distant-supervision setting.