Original title: Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation
Hasan Alkhder, Mohammad Abboush, Igor Tchappi, Ahmet Zengin +1 · September 2026
Most translation tools only give one output for a sentence. This hides other valid ways to say the same thing. This is a problem for low-resource dialects, which are regional language varieties with scarce training data. In dialects, words can differ in authenticity, formal tone, or stability. Researchers tackled this by treating translation as a decision space explored by autonomous agents. Each agent represents a different translation path running on a shared multilingual system. The researchers tested Turkish to Syrian Arabic using three agents: a direct zero-shot agent, an agent stabilized with lightweight fine-tuning, and an agent that routes through English. They found that differences between agents reveal useful behavioral signals instead of simple errors. Lightweight stabilization nearly doubled dialect marker usage from 0.2266 to 0.4988 and reduced structural instability. English pivoting pushed translations toward standard Arabic and caused compression, while the zero-shot agent showed the highest decision variance.
Imagine you're choosing what to wear for a day with changing weather: you might pack a jacket, carry an umbrella, or just wear shorts. Each choice leads to different comfort levels, and comparing them shows what works best for the day. In this paper, the different translation methods are like those packing choices, and the differences between what they 'wear' reveal which approach is most authentic for preserving local dialect flavor.
Speakers of low-resource dialects like Syrian Arabic often receive translations that lack local authenticity. This work shows how exploring multiple translation paths can identify and improve dialect-specific language, making communication more natural for these communities.
The abstract doesn't state explicit limitations, but the study is limited to Turkish-Syrian Arabic translation using dialogue data and specific behavioral metrics, without generalizing to other languages or real-world deployment.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Translation as a Decision Space: A Multi-Agent Perspective on Low-Resource Dialect Generation — Hasan Alkhder, Mohammad Abboush, Igor Tchappi et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.