Reproduce “Revisiting Projection-based Data Transfer for Cross-Lingual Named Entity Recognition in Low-Resource Languages” 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
Andrei Politov, Oleh Shkalikov, René Jäkel, Michael Färber
This research improves cross-lingual Named Entity Recognition for low-resource languages by developing a formalized candidate matching method for annotation projection and exploring refined word alignment techniques. The approach demonstrates performance gains over existing heuristics and, in several instances, surpasses direct multilingual model transfer, particularly through its NER model-based candidate extraction.
Key insights
- Formalizing annotation projection as a weighted bipartite matching problem provides a more robust and flexible framework than ad-hoc heuristics, generally improving performance across various low-resource languages.
- Employing a multilingual NER model to extract target candidate spans can effectively refine entity labels and, in some cases, yield better results than direct model transfer.
- The proposed target-to-target ("tgt2tgt") word alignment strategy, while conceptually promising, is highly dependent on the quality of the back-translation model and did not consistently outperform source-to-target alignment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.