# Word Translation Without Parallel Data

**Authors:** Alexis Conneau, Guillaume Lample, Marc'Aurelio Ranzato, Ludovic Denoyer, Hervé Jégou  
**arXiv:** [1710.04087](https://arxiv.org/abs/1710.04087)  
**Date:** 2017-10-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1710.04087v3/

## Key points

- Traditional methods for learning cross-lingual word embeddings rely on bilingual dictionaries or parallel corpora
- Recent studies have shown that character-level information can reduce the need for parallel data supervision
- Character-based methods still fall short compared to supervised approaches and are limited to language pairs with a common alphabet
- The authors propose a novel approach that builds a bilingual dictionary between two languages without using any parallel corpora
- Their model outperforms existing supervised methods on cross-lingual tasks for certain language pairs without utilizing character information
- Experiments demonstrate the effectiveness of their method for distant language pairs such as English-Russian or English-Chinese
- Experiments conducted on the low-resource language pair of English-Esperanto showcase potential impact in fully unsupervised machine translation
- Code, embeddings, and dictionaries developed by the authors are publicly available for further research and exploration

## Summary

In their paper titled "Word Translation Without Parallel Data," authors Alexis Conneau, Guillaume Lample, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou discuss state-of-the-art methods for learning cross-lingual word embeddings. These methods have traditionally relied on bilingual dictionaries or parallel corpora. However, recent studies have shown that the need for parallel data supervision can be reduced by incorporating character-level information. While these character-based methods have shown promising results, they still fall short compared to supervised approaches and are limited to language pairs that share a common alphabet. To address these limitations, the authors propose a novel approach that builds a bilingual dictionary between two languages without using any parallel corpora. This is achieved by aligning monolingual word embedding spaces in an unsupervised manner. Remarkably, the authors' model even outperforms existing supervised methods on cross-lingual tasks for certain language pairs without utilizing any character information. The experiments conducted demonstrate the effectiveness of their method for distant language pairs such as English-Russian or English-Chinese. Additionally, the authors describe experiments conducted on the low-resource language pair of English-Esperanto where only limited parallel data exists; their method showcases its potential impact in fully unsupervised machine translation. The code, embeddings, and dictionaries developed by the authors are publicly available for further research and exploration. Overall, this paper presents a significant advancement in cross-lingual word embedding learning by eliminating the reliance on parallel data and achieving impressive results across various language pairs.