Reproduce “FactMix: Using a Few Labeled In-domain Examples to Generalize to Cross-domain Named Entity Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Linyi Yang, Lifan Yuan, Leyang Cui, Wenyang Gao, Yue Zhang
FactMix introduces a data augmentation method that significantly enhances Named Entity Recognition (NER) models' ability to generalize to new domains using only a few labeled in-domain examples. The approach, which generates semi-factual examples by manipulating both entity and non-entity contexts, achieves up to a 14.19% absolute F1 score increase in challenging cross-domain few-shot settings compared to previous state-of-the-art methods.
Key insights
- Explicitly addressing and mitigating spurious correlations by manipulating non-entity contexts through 'context-level semi-facts' is crucial for robust cross-domain generalization in NER.
- A two-pronged rationale-centric data augmentation strategy, combining context and entity modifications, allows models to learn more causal and less domain-specific entity-context dependencies.
- A fully automatic and model-agnostic data augmentation framework can significantly improve few-shot NER performance across diverse domains without requiring any target-domain labels or human intervention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.