Using LLMs to translate Slovak
I have previously recommended specialized translation services for English-Slovak translation. LLMs like ChatGPT and Claude have however improved substantially in the meantime. They are now clearly superior to specialized translation services.
Why LLMs
LLMs have two key advantages:
- Steerability: You can guide the translation process by specifying terminology, style, and the balance between literal translation and fluency. You can also request edits or alternative translations easily.
- Cross-sentence fluency: LLMs understand context beyond individual sentences, resulting in more coherent translations overall.
Challenges with Slovak language
Slovak is a moderately difficult language for LLMs to master. Translation from Slovak to English is generally easier. The reverse direction is more challenging. Also, while many LLMs can produce understandable Slovak, achieving fluent, publishable quality is a higher bar. In my experience, only a few LLMs, notably ChatGPT and Claude, can consistently produce fluent Slovak translations.
Step-by-step guide
Here's a process I've found effective for obtaining high-quality Slovak translations using LLMs:
- Start with a basic translation request to get an initial draft.
- Add instructions to address article-wide issues like specific terminology.
- Ask the LLM to fix any mistranslations in style, word choice, and tone. This often results in a more fluent version, sometimes even surpassing the source text in fluency.
- Repeat the process with another LLM to get multiple unique drafts.
- Select the most appealing draft. I often find Claude's fluent version to be the best.
- Proofread the chosen draft, replacing problematic sentences with alternatives from the other drafts.
- If necessary, ask the LLM to provide multiple alternative translations for any remaining troublesome sentences.
This approach combines a few high-level operations to produce publishing-quality translations with minimal effort. It's significantly more convenient and effective than manually correcting output from dedicated translation services.