Machine Translation
Translation is the task that paid for NLP research for thirty years and keeps paying now. A model reads a sentence in one language and produces a sentence in another. Length varies. Word order varies. Some source words map to multiple target words and vice versa. Idioms refuse one-to-one mapping. "I miss you" in French is "tu me manques" — literally "you are lacking to me." No word-level alignment survives that. Machine translation is the task that forced NLP to invent encoder-decoders, attention, transformers, and eventually the whole LLM paradigm. Every step forward arrived because translation quality was measurable and the gap between human and machine was stubborn. This lesson skips the history lesson and teaches the working pipeline of 2026: pretrained multilingual encoder-decoder (NLLB-200 or mBART), subword tokenization, beam search, BLEU and chrF evaluation, and the handful of failure modes that still ship to production uncaught. MT pipeline: tokenize → encode → decode with attention → detokenize Modern MT is a transformer encoder-decoder trained on parallel text. The encoder reads the source in its language's tokenization. The decoder generates the target, one subword at a time, using the encoder's output via cross-attention (lesson 10). Decoding uses beam search to avoid the greedy-decoding trap. The output is detokenized, detruecased, and scored against a reference. Three operational choices drive real-world MT…
Machine Translation: Translation is the task that paid for NLP research for thirty years and keeps paying now. A model reads a sentence in one language and…
This free lesson is part of the AI Engineering from Scratch curriculum. Read the full explanation, run the lesson code, and verify the result in the interactive reader or from the repository source.
Browse the complete course catalog or open this lesson on GitHub.