Robust Cross-Lingual Transfer through Label Denoising

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2379 papers475 conceptsfrozen 2026-10-04

The central challenge in cross-lingual Natural Language Processing (NLP) is a data paradox: high-resource languages like English have an abundance of labeled data, while thousands of other languages have almost none. To bridge this gap, researchers typically use "label projection"—translating English data into a target language and assuming the labels (like "Positive Sentiment" or "Person Name") remain valid.

However, this process is inherently messy. Machine translation (MT) is not a perfect mirror; it introduces "noise" that can derail a model's learning. This research direction focuses on identifying, filtering, and mitigating this noise to build robust models that can learn from imperfect, "silver-standard" data.

The Anatomy of Cross-Lingual Noise

Why does cross-lingual transfer fail? The papers in this field identify several recurring failure modes.

First, there is semantic polarity flipping. As Qiang Chen and colleagues demonstrate in Learning to Adapt Credible Knowledge in Cross-lingual Sentiment Analysis, MT systems can inadvertently flip the sentiment of a sentence—turning a negative review into a positive one—in roughly 10% of cases. If a model treats these translated labels as ground truth, it internalizes incorrect logic.

Second, there is structural and alignment divergence. Languages do not map one-to-one. Imed Zitouni and Radu Florian highlight in Mention Detection Crossing the Language Barrier that complex many-to-many word alignments make it difficult to project token-level labels (like Named Entities) accurately. A single English word might be translated into a multi-word phrase in Turkish, leaving the model confused about where an entity begins or ends.

Finally, there is conceptual drift. In UOM-FJ at the NTCIR-15 SHINRA2020-ML Task, Hiyori Yoshikawa notes that even Wikipedia's inter-language links can be noisy; a specific entity in one language might link to a much broader conceptual category in another, leading to misaligned training signals.

From Absolute Truth to Soft Suggestions

To combat this noise, researchers have moved away from treating projected labels as "hard" facts. Instead, they treat them as "soft" constraints or suggestions.

A landmark approach in this area is Posterior Regularization. In Cross-Lingual Discriminative Learning of Sequence Models with Posterior Regularization, Kuzman Ganchev and Dipanjan Das argue that models should be allowed to deviate from projected labels if those labels violate the target language's internal logic. By using regularization, the model can resolve contradictions using target-language context. Similarly, Andre Martins applied this to coreference resolution in Transferring Coreference Resolvers with Posterior Regularization, using global constraints to fix broken entity chains caused by alignment errors.

Filtering the Signal from the Noise

Another dominant strategy is active denoising—identifying and discarding the "bad" translations before the model ever sees them.

Enforcing Consistency

The most recent wave of research focuses on Consistency Training. The intuition is simple: if you translate a sentence into a target language and then back to the source, or if you view the same data through different "dropout" masks, the model's predictions should remain stable.

Ran Zhou and Lidong Bing demonstrate this in ConNER, where enforcing consistency across different views of unlabeled target-language data helps the model ignore the "noise" of unnatural phrasing. Similarly, the CoLaDa framework by Qianhui Wu (CoLaDa: A Collaborative Label Denoising Framework for Cross-lingual Named Entity Recognition) uses iterative model collaboration to "clean" noisy labels by looking for neighborhood-based instance consistency.

By shifting the focus from perfect translation to intelligent denoising, this research direction enables AI to expand into low-resource languages without the prohibitive cost of manual annotation.

Papers in this version

Showing 60 of 379 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 199 papers
  2. Version 1 · 2026-10-04 · 379 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_cross_lingual_projection_noise_78dea33b,
  title = {cross-lingual projection noise},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-projection-noise/rdr_78dea33b414446399c5a3c2a88e963db},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_cross_lingual_transfer_through_la_78dea33b_v1,
  title = {Robust Cross-Lingual Transfer through Label Denoising},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-cross-lingual-transfer-through-label-denoising/rdr_78dea33b414446399c5a3c2a88e963db/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}