= Machine translation
<Machine translation> uses a computational system to produce text in one human language from text in another. It must choose an output sequence that preserves relevant meaning while fitting the target language, rather than merely replacing each word independently. A <transformer model> is one architecture used for this task.<Attention Is All You Need by Ashish Vaswani and others>[§3, encoder-decoder sequence model; §6.1, machine translation evaluation]{ref}
A probabilistic formulation assigns <conditional probabilities> $p(y\mid x)$ to candidate output sequences $y$ for input $x$. A decoder searches for a suitable sequence, often using model probability together with additional search or length rules. For a deliberately simplified finite example, suppose three complete candidates have probabilities $0.55$, $0.30$ and $0.15$. Choosing the most probable gives the first candidate, but this is a model decision rather than a proof that its meaning is correct.
Context explains why word-by-word replacement is insufficient. The intended translation of an ambiguous word can depend on the surrounding sentence, and target word order can differ from source word order. A system can therefore produce a fluent sentence that omits or alters information; fluency and preservation of meaning require separate attention.
Training and evaluation also answer different questions. A <neural language model> can supply probabilities for sequences, while <cross-entropy> measures how much probability is assigned to observed text. <BLEU> instead compares output sequences with reference translations through matching <n-grams> and a length penalty. Reference overlap is useful for repeated comparisons, but alternative valid phrasing and context limit its interpretation as an assessment of any single sentence.<BLEU: a Method for Automatic Evaluation of Machine Translation by Kishore Papineni and others>[§1.1, motivation; §2.1.1 and footnote 4, corpus evaluation versus individual sentences]{ref}
Back to article page
