Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Constrained Decoding for Cross-lingual Label Projection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Codec, a constrained-decoding method for label projection that translates text without markers first and then inserts span markers via branch-and-bound search, enabling both translate-train and translate-test for NER and event argument extraction across 20 languages.
Problem
Marker-based label projection (EasyProject) degrades translation quality because span markers are injected into the MT input, and it cannot be used for translate-test; alignment-based methods are less accurate.
Approach
Translate the source without markers to get a template, then use constrained decoding (branch-and-bound with a heuristic lower bound and opening-marker position pruning) conditioned on the marked source to insert markers into the fixed template, with span-level re-ranking.
Key results
- On MasakhaNER2.0 translate-test, Codec averages 70.4 F1 versus 65.8 for Awes-align and 54.5 for FT-En across 18 languages.
- In translate-train, Codec averages 67.1 F1 versus 64.9 for EasyProject on MasakhaNER2.0.
- For Bambara, Codec with heuristic bound is about 60x faster than exact search while losing 0.6 F1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoDeC instead of EasyProject for translate-train label projection | mDeBERTa-v3 NER · MasakhaNER2.0 test (AVG) | F1: 64.9 -> 67.1 | Helped |
| Use translate-test CoDeC rather than translate-train CoDeC | mDeBERTa-v3 NER · MasakhaNER2.0 test (AVG) | F1: 67.1 -> 70.4 | Helped |
| Use CoDeC instead of Awes-align for translate-train label projection | mT5-large EAE · ACE-2005 test (AVG) | F1: 52.8 -> 53.8 | Helped |
| Fine-tune the decoding MT on synthetic data with event labels constructed by string matching (ablation) | CoDeC EAE · ACE-2005 translate-test, Arabic with GMT template | F1: 49.1 -> 53.9 | Helped |
| Fine-tune the decoding MT on synthetic data with event labels constructed by alignment (ablation) | CoDeC EAE · ACE-2005 translate-test, Chinese with GMT template | F1: 54.7 -> 56.8 | Helped |
| Return only the highest-probability hypothesis instead of reranking candidates (ablation) | CoDeC · MasakhaNER2.0 translate-dev, five-language sample | F1 | Hurt |
| Use the heuristic lower bound with δ=3 instead of the exact lower bound (ablation) | CoDeC · MasakhaNER2.0 translate-dev, five-language sample | F1 | No clear effect |
| Prune unlikely opening-marker positions (ablation) | CoDeC with δ=3 · MasakhaNER2.0 translate-dev, five-language sample | F1: 0.7 absolute F1 | Mixed |
| Use NLLB-1.3B instead of NLLB-600M for template translation (ablation) | CoDeC NER · MasakhaNER2.0 translate-test (AVG) | F1: 68.6 -> 70.6 | Helped |
| Use NLLB-3.3B instead of NLLB-600M for constrained decoding with the same templates (ablation) | CoDeC NER · MasakhaNER2.0 translate-test (AVG) | F1: 70.4, 70.1, 70.1 | No clear effect |
| Use translate-test CoDeC instead of constrained-space beam search with beam size 16 | mDeBERTa-v3 NER · MasakhaNER2.0 test (AVG) | F1: 62.3 -> 70.4 | Helped |
Surprising findings
- Translate-test with label projection substantially outperforms translate-train on African-language NER, with FT-En even best on Chichewa and Kiswahili.
- Pruning opening-marker positions can boost F1 in several languages, suggesting the MT model assigns high probability to bad marker hypotheses.
Limitations
- Re-ranking fails to select the correct hypothesis in over 60% of Chinese error cases, per manual analysis.
- Translate-test with event arguments requires an extra projection pass; EAE evaluated only in translate-train in main results.
- Decoding MT for EAE needed custom synthetic fine-tuning; NLLB tokenizer lacks some Chinese characters.
Experimental setup
Datasets: CoNLL03 MasakhaNER2.0 ACE-2005 WikiANN NLLB training data VLSP 2016 NER MSRA-NER CoNLL-2002/2003
Benchmarks: MasakhaNER2.0 ACE-2005
Metrics: F1
Baselines: EasyProject Awes-align FT-En GPT-4 Constrained-Space Beam Search M2M-100
Builds on: EasyProject Awesome-align X-Gear NLLB
Closest papers by meaning
- FILTER: An Enhanced Fusion Method for Cross-lingual Language Understanding (AAAI 2021 · similarity 0.88)
- Cross-Lingual Natural Language Generation via Pre-Training (AAAI 2020 · similarity 0.88)
- Cross-Lingual Pre-Training Based Transfer for Zero-Shot Neural Machine Translation (AAAI 2020 · similarity 0.88)
- Evaluating the Cross-Lingual Effectiveness of Massively Multilingual Neural Machine Translation (AAAI 2020 · similarity 0.88)
- Unsupervised Interlingual Semantic Representations from Sentence Embeddings for Zero-Shot Cross-Lingual Transfer (AAAI 2020 · similarity 0.87)
- CL2CM: Improving Cross-Lingual Cross-Modal Retrieval via Cross-Lingual Knowledge Transfer (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.