3.36 kB
3 files
Updated about 2 months ago
Name
Size
README.md2.36 kB
xet
results.csv368 Bytes
xet
results_scored.csv633 Bytes
xet
README.md

Somali Is Not a Second-Class Language: Translation Errors a Native Speaker Can See

1. The blind spot (Question 1)

Somali is an under-resourced language, and we are no longer accepting being second-class citizens. When I try to chat with AI, specifically ChatGPT, the context and the language are out of scope: too many errors in translation and conversation. I asked it to explain what a supervised model is in Somali, and the context was poor and misleading; someone who is a native Somali speaker cannot even read it and connect. So the Somali language should be improved.

2. Model tested

  • Model: qvac/TranslatePsy-AfriSLM-4B
  • Why I chose it: It's one of the few models built specifically for African languages, including Somali, and it claims to beat much larger models. As a native speaker, I wanted to see if that holds up on everyday phrases.

3. What I tested and found

I gave the model English sentences, it translated them into Somali, and I judged each translation as a native Somali speaker.

English Model's Somali My correct Somali What went wrong
How are you today? Salaan kadib sidee tahay maanta? Iska waran maanta? / Sidee tahay maanta? "Salaan kadib" means "after greeting". It is misleading and not included in "How are you today?"
welcome with open hands soo dhawoow gacmo furan Ku soo dhawoow gacmo furan "Ku" is missing. It is needed to indicate welcoming someone.
Who do you think you are Yaa aad u maleyneysaa in aad tahay? Yaad u malaynaysaa in aad adigu tahay? The context is wrong and out of the topic. It is philosophy and grammatically hard; I guess only a good model or a native Somali speaker can translate it.

What this shows: The model got all three sentences wrong: it added words, dropped a word, and missed an idiom's meaning. It can produce Somali-looking sentences but doesn't sound like a native speaker. Three sentences is a small test, but it shows the output still needs a native speaker to check it.

4. How to fix it (Question 3)

Most of its training data was machine-generated, so it learned those mistakes. More real Somali text written or reviewed by native speakers, plus human testing before release, would help.

Files

  • results_scored.csv: the model's translations with my judgments
  • the notebook (.ipynb): the code I ran
Total size
3.36 kB
Files
3
Last updated
Sep 29
Pre-warmed CDN
US EU US EU

Contributors