The Normalization of AI Translation: Principles of LLM-Based Translation and Structural Differences from NMT

AI translation has become a routine communication tool, integrated into major platforms and devices such as Netflix, YouTube, and Samsung Galaxy. Translators are adopting workflows where they use Claude or Gemini instead of CAT tools to handle short sentences or urgent tasks, followed by human review.

How LLM Translation Works

Current mainstream LLM-based translation is the result of implicitly acquiring cross-lingual correspondences through vast multilingual text, without separate translation training. LLMs adopt a Decoder-Only structure, processing source and target texts as a single continuous context, and calculating relationships between tokens via self-attention. This contrasts with the Encoder-Decoder structure of existing NMT (Neural Machine Translation), where the encoder and decoder are separated.

Practical Differences Between NMT and LLM

The two approaches show clear differences in processing units and speed. LLMs can process entire documents or conversations and immediately reflect instructions via prompts, but consistency may waver in long documents. Conversely, NMT makes it easier to enforce consistency through sentence-level processing. In terms of speed, LLMs are 10 to 100 times slower than NMT, so NMT remains preferred in environments where latency is critical. Ultimately, AI translation is closer to probabilistic guessing based on learned patterns rather than perfect translation.