# Comparing Formulaic Language in Human and Machine Translation: Insight from a Parliamentary Corpus

**Authors:** Yves Bestgen  
**arXiv:** [2206.10919](https://arxiv.org/abs/2206.10919)  
**Date:** 2022-06-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.10919v1/

## Key points

- Study aims to replicate previous research comparing neural machine translations to human translations
- Previous study found that neural machine translations have more formulaic sequences with high-frequency words, but fewer with rare words compared to human translations
- Researchers used a parliamentary corpus to replicate the findings
- Corpus was translated from French to English using DeepL, Google Translate, and Microsoft Translator
- Results confirmed previous observations but with less pronounced differences
- Suggests that using text genres resulting in more literal translations (e.g., parliamentary corpora) is preferable for comparing human and machine translations
- Google translations had fewer highly collocational bigrams than DeepL and Microsoft translations
- Findings provide insights into differences between neural machine translation systems and emphasize the importance of considering text genre when evaluating translation quality.

## Summary

This study aims to replicate previous research that compared neural machine translations to human translations. The previous study found that neural machine translations contain more strongly-associated formulaic sequences made of high-frequency words, but fewer strongly-associated formulaic sequences made of rare words, compared to human translations. In this study, the researchers used a parliamentary corpus to see if the findings could be replicated. The text in the corpus was translated from French to English using three well-known neural machine translation systems: DeepL, Google Translate, and Microsoft Translator. The results confirmed the observations from the news corpus but with less pronounced differences. This suggests that using text genres that typically result in more literal translations, such as parliamentary corpora, may be preferable when comparing human and machine translations. Additionally, the study found that Google translations contained fewer highly collocational bigrams than DeepL and Microsoft translations. These findings provide insights into the differences between neural machine translation systems and highlight the importance of considering text genre when evaluating translation quality.