Research map / Graph learning / Graph self-supervised learning
Analysis of Joint Multilingual Sentence Representations and Semantic K-Nearest Neighbor Graphs
Holger Schwenk, Douwe Kiela, Matthijs Douze
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| joint multilingual encoder instead of bilingual encoders (EN-FR) | Joint sentence encoder · UN corpus EN-FR | P@1: 48.90 -> 83.30 | Improved |
| joint multilingual encoder instead of bilingual encoders (EN-ES) | Joint sentence encoder · UN corpus EN-ES | P@1: 54.94 -> 85.40 | Improved |
| compress k-nn index with OPQ and IVF clustering | FAISS index · 566M-sentence graph | storage: 2.4 TB -> 22 GB | Improved |
| PCA dimension reduction instead of FAISS compression | PCA · 566M-sentence index | search error | Hurt |
| encode in-domain Europarl-trained model on short informal Tatoeba sentences | Joint sentence encoder · Tatoeba vs Europarl | similarity error: 0.84 -> higher | Hurt |
| zero-shot Russian on UN corpus versus Tatoeba | Joint sentence encoder · EN-RU | error: 65.84 vs 49.2 | Mixed |
| zero-shot to unseen related Slavic languages | Joint sentence encoder · Tatoeba zero-shot | similarity error (%): 36.4 | Improved |
| zero-shot to Norwegian Bokmål and Galician | Joint sentence encoder · Tatoeba zero-shot | similarity error (%): 30; 20.1 | Improved |
| zero-shot to unrelated Latin-script languages (Turkish, Vietnamese, Indonesian, Tagalog) (ablation) | Joint sentence encoder · Tatoeba zero-shot | similarity error (%): above 95 | Hurt |
| zero-shot to Esperanto (partial overlap with European languages) (ablation) | Joint sentence encoder · Tatoeba zero-shot | similarity error (%): 71 | Mixed |
| mixing up to seven languages in one sentence (code-switching) | Joint sentence encoder · code-switching queries | nearest neighbor correct | Improved |
Closest papers by meaning
- A Comparison of Architectures and Pretraining Methods for Contextualized Multilingual Word Embeddings (AAAI 2020 · similarity 0.88)
- Modeling Sequential Sentence Relation to Improve Cross-lingual Dense Retrieval (ICLR 2023 · similarity 0.88)
- Narrowing the Gap between Supervised and Unsupervised Sentence Representation Learning with Large Language Model (AAAI 2024 · similarity 0.86)
- Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models (NeurIPS 2025 · similarity 0.86)
- MMTEB: Massive Multilingual Text Embedding Benchmark (ICLR 2025 · similarity 0.86)
- MKGL: Mastery of a Three-Word Language (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.
