# Bilex Rx: Lexical Data Augmentation for Massively Multilingual Machine Translation

**Authors:** Alex Jones, Isaac Caswell, Ishank Saxena, Orhan Firat  
**arXiv:** [2303.15265](https://arxiv.org/abs/2303.15265)  
**Date:** 2023-03-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.15265v1/

## Key points

- Neural machine translation (NMT) has made significant progress, especially with the development of Unsupervised Machine Translation (UNMT)
- UNMT allows for high-quality translation using only monolingual text data
- NMT models struggle with accurately translating common nouns
- Bilingual lexica can be used to address this issue
- Experiments on 200-language translation models showed that augmenting lexical data led to substantial performance improvements in unsupervised translation
- Different families of data augmentation techniques yielded similar improvements and could be combined for greater gains
- Carefully curated lexica were found to be more important than larger but noisier ones, especially when used with larger models
- Multilingual lexicon data was compared to human-translated parallel data, and its efficacy was evaluated
- GATITOS, a new multilingual lexicon for 26 low-resource languages, outperformed other lexica in the experiments and is open-source
- Examples of mistakes commonly made by NMT models were discussed, along with how bilingual lexicon entries can help mitigate these errors
- Details about the datasets used in the experiments, including monolingual and parallel data sources, were provided
- Augmenting training data with bilingual lexicon information significantly improves machine translation performance for low-resource and unsupervised languages
- The GATITOS dataset achieved average gains of around +8 CHRF on unsupervised languages
- Future research may focus on leveraging bilingual lexica to prompt foundation models, particularly for low-resource languages

## Summary

Neural machine translation (NMT) has made significant progress in recent years, particularly with the development of Unsupervised Machine Translation (UNMT), which allows for high-quality translation using only monolingual text data. However, these models still struggle with certain aspects of translation that are easy for humans, such as accurately translating common nouns. To address this issue, this work explores the use of bilingual lexica as a valuable and abundant resource. The researchers conducted experiments on 200-language translation models trained on web-crawled text to test the effectiveness of bilingual lexica in real-world scenarios. Several key findings emerged from their study: by augmenting lexical data, they achieved substantial performance improvements in unsupervised translation; different families of data augmentation techniques yielded similar improvements and could be combined for even greater gains; carefully curated lexica were found to be more important than larger but noisier ones, especially when used with larger models; and the efficacy of multilingual lexicon data was compared to human-translated parallel data. Additionally, the researchers introduced GATITOS, a new multilingual lexicon for 26 low-resource languages that outperformed other lexica in their experiments. This dataset is open-source and can be accessed at https://github.com/google-research/url-nlp/tree/main/gatitos. The paper also discussed examples of mistakes commonly made by NMT models and how bilingual lexicon entries can help mitigate these errors. Furthermore, it provided details about the datasets used in the experiments, including monolingual and parallel data sources. In conclusion, this paper demonstrated that augmenting training data with bilingual lexicon information significantly improves machine translation performance for low-resource and unsupervised languages. The GATITOS dataset proved particularly effective in achieving average gains of around +8 CHRF on unsupervised languages. Future research may focus on leveraging bilingual lexica to prompt foundation models, especially for low-resource languages.