Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Multilingual Alignment of Contextual Word Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune multilingual BERT for contextual word alignment using parallel data | multilingual BERT · XNLI test (zero-shot), Average | accuracy: 72.4 -> 74.7 | Helped |
| Use fine-tuning-based alignment instead of rotation-based word alignment | multilingual BERT · XNLI test (zero-shot), Average | accuracy: 72.2 -> 74.7 | Helped |
| Fine-tune for alignment rather than use the base model | multilingual BERT · XNLI test (zero-shot), Bulgarian | accuracy: 68.7 -> 73.4 | Helped |
| Fine-tune for alignment rather than use the base model | multilingual BERT · XNLI test (zero-shot), Greek | accuracy: 67.0 -> 71.4 | Helped |
| Fine-tune for contextual alignment instead of using sentence-augmented aligned fastText | contextual word retrieval · Word retrieval, contextual, Average | accuracy: 45.1 -> 50.3 | Helped |
| Fine-tune for contextual alignment rather than align BERT with a rotation | contextual word retrieval · Word retrieval, contextual, Average | accuracy: 30.0 -> 50.3 | Helped |
| Fine-tune BERT for alignment rather than use sentence-augmented aligned fastText | non-contextual word retrieval · Word retrieval, non-contextual, Average | accuracy: 64.8 -> 65.9 | Helped |
| Use 10K parallel sentences for alignment rather than no alignment data (ablation) | multilingual BERT · XNLI test (zero-shot), Average | accuracy: 72.4 -> 73.2 | Helped |
| Increase alignment data from 10K to 50K parallel sentences (ablation) | multilingual BERT · XNLI test (zero-shot), Average | accuracy: 73.2 -> 74.3 | Helped |
| Increase alignment data from 50K to 250K parallel sentences (ablation) | multilingual BERT · XNLI test (zero-shot), Average | accuracy: 74.3 -> 74.7 | Helped |
| Fine-tune for alignment using 20K parallel sentences per language | multilingual BERT · XNLI test (zero-shot), Average; additional languages | accuracy: 70.1 -> 71.1 | Helped |
| Fine-tune for alignment rather than use the base model | multilingual BERT · Non-contextual word retrieval, open-class POS average | accuracy: 0.54 -> 0.89 | Helped |
Closest papers by meaning
- Cross-lingual Alignment vs Joint Training: A Comparative Study and A Simple Unified Framework (ICLR 2020 · similarity 0.90)
- A Comparison of Architectures and Pretraining Methods for Contextualized Multilingual Word Embeddings (AAAI 2020 · similarity 0.90)
- Normalization of Language Embeddings for Cross-Lingual Alignment (ICLR 2022 · similarity 0.89)
- Unsupervised Hyper-alignment for Multilingual Word Embeddings (ICLR 2019 · similarity 0.89)
- MMTEB: Massive Multilingual Text Embedding Benchmark (ICLR 2025 · similarity 0.86)
- On Affine Homotopy between Language Encoders (NeurIPS 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.