Research map / Label-efficient and robust learning / Recognition and long-tail
Comparing Unsupervised Word Translation Methods Step by Step
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 |
|---|---|---|---|
| use vanilla GAN distribution matching instead of Gromov-Wasserstein seed (same Procrustes refinement) | Procrustes refinement · MUSE et-en/lv-en/vi-en | P@1: 1.1 -> 27.5 (et-en) | Helped |
| use ICP or Alvarez-Melis GW instead of vanilla GAN for seed induction | Procrustes refinement · MUSE six language pairs | P@1: 27.5 -> 0.1 (et-en) | Hurt |
| add stochastic dictionary induction to MUSE with cosine model selection (vs Procrustes) (ablation) | C-MUSE · MUSE average P@1 | P@1: 45.4 -> 52.1 | Helped |
| replace Artetxe GW initialization with vanilla GAN initialization under stochastic dictionary induction | Artetxe et al. (2018) · MUSE average P@1 | P@1: 42.3 -> 52.1 | Helped |
| add Procrustes refinement to vanilla GAN seed (ablation) | MUSE vanilla GAN · et-en | P@1: 6.4 -> 27.5 | Helped |
| apply unit length normalization to embeddings (ablation) | MUSE vanilla GAN · six language pairs | failed runs | Hurt |
| inject 25% random vectors (ablation) | MUSE vanilla GAN · a few languages | P@1 | No clear effect |
| wider discriminators (over-parameterization) (ablation) | MUSE vanilla GAN · six language pairs | stability and P@1 | Mixed |
| larger batch size or changed learning rate (ablation) | MUSE vanilla GAN · six language pairs | P@1 | Hurt |
| use Wasserstein GAN or CT-GAN instead of vanilla GAN | MUSE vanilla GAN · UBDI language pairs | P@1 and stability | Hurt |
| use discriminator loss vs cosine similarity of nearest neighbors for model selection | GAN restarts · UBDI | correlation with P@1 | Helped |
Closest papers by meaning
- Unsupervised Cross-Modal Alignment of Speech and Text Embedding Spaces (NeurIPS 2018 · similarity 0.90)
- Normalization of Language Embeddings for Cross-Lingual Alignment (ICLR 2022 · similarity 0.89)
- Word translation without parallel data (ICLR 2018 · similarity 0.89)
- Unsupervised Hyper-alignment for Multilingual Word Embeddings (ICLR 2019 · similarity 0.88)
- CL2CM: Improving Cross-Lingual Cross-Modal Retrieval via Cross-Lingual Knowledge Transfer (AAAI 2024 · similarity 0.86)
- Unified Lexical Representation for Interpretable Visual-Language Alignment (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.