# Improving Supervised Bilingual Mapping of Word Embeddings

**Authors:** Armand Joulin, Piotr Bojanowski, Tomas Mikolov, Edouard Grave  
**arXiv:** [1804.07745](https://arxiv.org/abs/1804.07745)  
**Date:** 2018-04-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1804.07745v1/

## Key points

- Authors: Armand Joulin, Piotr Bojanowski, Tomas Mikolov, Edouard Grave
- Focus: Alignment of continuous word representations across different languages
- Precision achieved using small bilingual lexicon for training
- Hubness problem in linear transformation process for retrieval tasks like word translation
- Proposal to use retrieval criterion instead of square loss for learning mapping
- Alternative loss function yields state-of-the-art results in word translation tasks
- Significant improvements observed for distant language pairs such as English-Chinese
- Novel approach to enhancing supervised bilingual mapping by addressing hubness and retrieval task challenges
- Effectiveness of incorporating retrieval criterion into learning process for improved cross-lingual performance

## Summary

In their paper titled "Improving Supervised Bilingual Mapping of Word Embeddings," authors Armand Joulin, Piotr Bojanowski, Tomas Mikolov, and Edouard Grave explore the alignment of continuous word representations across different languages. They highlight the precision achieved in aligning these representations by using a small bilingual lexicon for training. The authors note that the linear transformation process can lead to the hubness problem when applied to retrieval tasks like word translation. To mitigate this issue, they propose utilizing a retrieval criterion instead of the square loss for learning the mapping. Through their evaluation on word translation tasks, they demonstrate that this alternative loss function yields state-of-the-art results. Particularly noteworthy are the significant improvements observed for distant language pairs such as English-Chinese. This study sheds light on a novel approach to enhancing supervised bilingual mapping of word embeddings by addressing challenges related to hubness and retrieval tasks. The findings underscore the effectiveness of incorporating a retrieval criterion into the learning process, ultimately leading to improved performance in cross-lingual applications.