Celebrating World Translation Day: Why we should use machines for speed, but humans for accuracy.

Advances in AI and natural language processing have made translation tools faster and more efficient. However, how well are they supporting marginalised languages? To mark World Translation Day, we conducted an informal test by translating five Kinyarwanda news articles into English using various platforms, with a native speaker reviewing accuracy. The results were mixed—some platforms performed better, but all required human review for accuracy.

Most surprising translation of the experiment? Microsoft Translate translating ‘President Kagame’ to ‘President Obama’.

This aligns with the broader consensus that current translation technologies fall short in serving marginalised languages - further deepening social and economic divides by limiting access to critical resources like education, employment, and civic participation for speakers of these languages.

Informal Experiment Results: Kinyarwanda-to-English Translation

Google Translate	Microsoft Translator
Accuracy: 70%	Accuracy: 30%
Performance: Reasonable with basic phrases but struggled with sentence structure and cultural nuances.	Performance: Frequent mistranslations, especially with names and cultural references.
Example text: "Basketball continues to provide entertainment for the locals."	Example text: "President Kagame held a meeting with the Secretary-General of the EAC."
Google Translate: "The game of basketball is famous and entertained many."	Microsoft Translator: "President Obama held a meeting with the Secretary-General of the EAC."

Phrase	Lokalise
Accuracy: 40%	Accuracy: 40%
Performance: Produced fragmented translations with awkward phrasing.	Performance: Missed key details, resulting in incomplete translations.
Example text: "Basketball continues to provide entertainment for the locals."	Example text: "Basketball continues to provide entertainment for the locals."
Phrase: "Basketball brings fun to the heart of the area people."	Lokalise: "Basketball is great many people watch."

AssemblyAI: Does not support Kinyarwanda translation.

Amazon Translate: Does not support Kinyarwanda

Key Takeaways

based on our informal Kinyarwanda-to-English Translation test

• Best Platform: Google Translate, with 70% accuracy, still needs human review for complex content.
• Worst Platform: Microsoft Translator, with shocking errors like confusing President Kagame with President Obama.
• Unsupported Platforms: Amazon Translate, AssemblyAI and DeepL do not support Kinyarwanda, and others like Phrase and Lokalise struggled with translation quality.

For INGOs working with marginalised communities, accurate translation is crucial—not only for effective communication but also for preserving the integrity of their research and outreach efforts. A hybrid approach, combining technology with human review, remains the best option for ensuring both speed and accuracy.

Here are some guiding principles for INGOs to achieve rapid and accurate translations of research projects:

1. Use machine translation for speed, but have human translators review for accuracy.
2. Run a simple experiment (like the one above) to identify the most effective platform for your required languages.
3. Use humans to simplify complex text before machine translation to reduce errors.
4. Implement a two-step human review process for cultural and contextual accuracy.
5. Prioritise human translation for sensitive content.
6. Crowdsource reviews from native speakers to ensure cultural relevance.

As we continue to explore how technology can better serve the translation needs of marginalised languages, more experiments are on the horizon. While we may not have billion-dollar budgets to develop these technologies ourselves, the question remains: how can tech better support the translation process in the INGO sector? Stay tuned as we dive deeper into finding solutions!

Want to hear more about our experiments? We would love to hear from you! Email hello@hereiamstudio.com