Just Use XML: Revisiting Joint Translation and Label Projection

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

LabelPigeon is a new framework that jointly performs machine translation and label projection using XML tags, challenging the prior assumption that combining these tasks degrades translation quality. The framework, which fine-tunes the NLLB-200 3.3B model on a modified Salesforce Localization XML MT dataset, demonstrates improved translation quality across 203 languages and superior label projection accuracy. LabelPigeon outperforms baselines like Awesome-align, Gemma 3 27B IT, and EasyProject in direct label projection evaluations on XQuAD and MLQA datasets, achieving up to +39.9 F1 score improvement in downstream tasks like Named Entity Recognition (NER). It also shows consistent gains in cross-lingual transfer across 27 languages and three downstream NLP tasks, including Question Answering (QA) and Coreference Resolution (CR), without incurring additional computational overhead at inference.

Key takeaway

Research Scientists developing multilingual NLP applications should reconsider multi-stage pipelines for label projection. By adopting LabelPigeon's XML-tag-based joint translation and label projection, you can achieve superior cross-lingual transfer performance, particularly for span-level tasks like NER, while simultaneously improving translation quality and reducing engineering complexity. Focus on fine-tuning with high-quality, XML-tagged parallel data to maximize these benefits.

Key insights

Joint translation and label projection with XML tags can improve both translation quality and label transfer.

Principles

• XML tags enable direct correspondence for span labels.
• Fine-tuning improves translation quality and label projection.
• Less idiomatic translations do not imply substantial quality loss.

Method

Fine-tune a base translation model (NLLB-200 3.3B) on XML-tagged parallel corpora, then use an off-the-shelf XML parser to extract tags after translation.

In practice

• Use XML tags for span-level annotations.
• Fine-tune on high-quality, XML-tagged parallel data.
• Prioritize three high-resource languages for fine-tuning.

Topics

• Label Projection
• Joint Translation
• XML Tagging
• Cross-lingual Transfer
• Named Entity Recognition

Code references

• thennal10/LabelPigeon

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Generating Concept Lexicalizations via Dictionary-Based Cross-Lingual Sense Projection — A dictionary-augmented projection method improves cross-lingual WordNet expansion precision with minimal resources.
• CzechDocs: A Multiway Parallel Dataset of Formatted Documents for Minority Languages in Czechia — A new multiway parallel dataset facilitates evaluating machine translation systems' ability to preserve document formatting, particularly for minority languages.
• NLP Token Classification : NER, POS Tagging and Chunking — Token classification is fundamental for NLP, enabling machines to deeply understand text through NER, POS tagging, and chunking.
• KIT's Submission to Cross-Lingual Voice Cloning in IWSLT 2026 — Explicit language tags and lexical matching significantly improve cross-lingual voice cloning performance and reduce accent leakage.
• CorPipe at CRAC 2026: Empty Nodes and Cross-Lingual Transfer in Multilingual Coreference Resolution — CorPipe 26 integrates empty node prediction with mention and coreference resolution for improved multilingual performance.
• TantraTagger: A Benchmark Dataset for Tantrayukti-Based Discourse Structure Labelling in Sanskrit Śāstra Texts — TantraTagger is a benchmark dataset for discourse structure labelling in Sanskrit Śāstra Texts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.