STREAMING UZBEK–RUSSIAN TEXT TRANSLATION WITH LATENCY-CONTROLLED TRANSFORMERS

Abstract

In this paper, we present a streaming Uzbek–Russian translation system based on a latency-controlled Transformer. The model combines wait-k scheduling, monotonic multi-head attention, and a lightweight look-ahead mechanism. To address the morphological complexity of Uzbek, we integrate morphology-aware tokenization and dynamic read/write policies. Our approach improves BLEU and chrF scores while reducing Average Lagging, and it remains robust in code-switching scenarios such as conversational text and chat data.