Machine Translation

Machine translation (MT) is the use of software to translate text or speech from one language into another without human intervention. It is one of the oldest and most visible applications of artificial intelligence to language, and its history reads like a map of the field's own paradigm shifts: from hand-written rules to statistical learning and, finally, to the neural networks that now power large language models.

The problem is harder than it looks. Translating is not swapping words but preserving meaning, register and context across linguistic systems that rarely map one to one. That difficulty explains why MT has needed three generations of technology to approach human quality.

From rules to statistics

The earliest systems, rule-based machine translation (RBMT), dominated the field's opening decades. Linguists hand-coded bilingual dictionaries and grammatical rules; the output was predictable but rigid, unable to absorb the ambiguity and variation of real language. In the 2000s the paradigm shifted to statistical machine translation (SMT): instead of rules, systems learned from large bilingual corpora which translations were most probable. The word-alignment models developed at IBM and, above all, the phrase-based approach formalized by Philipp Koehn, Franz Och and Daniel Marcu in 2003 set the industry standard for more than a decade.

The neural era: from the Transformer to LLMs

The decisive leap came with neural machine translation (NMT). In 2014, Ilya Sutskever, Oriol Vinyals and Quoc Le proposed the encoder-decoder (seq2seq) architecture, which compresses the source sentence into a vector and rebuilds it in the target language. That same year, Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio introduced the attention mechanism, letting the model look selectively at the relevant source words and solving the long-sentence bottleneck. In 2017, Ashish Vaswani and colleagues at Google pushed the idea to its conclusion with the Transformer («Attention Is All You Need»), which drops recurrence and processes the sentence in parallel; it reached 28.4 BLEU on the WMT 2014 English-German task and became the foundation of almost all MT that followed. Today, multilingual large language models translate with remarkable quality, often with no training specific to the language pair; in the WMT evaluation campaigns of 2023 and 2024, general-purpose systems topped many of the language pairs assessed, in several cases beating specialized neural systems.

How it is measured: BLEU and its limits

Scoring a translation automatically is a problem in its own right. The historical benchmark is BLEU, proposed by Kishore Papineni and colleagues in 2002: it compares a system's output against one or more human translations and computes the overlap of n-grams (sequences of one to four words), with a brevity penalty. It is fast, cheap and language-independent, but it measures surface word overlap, not adequacy or meaning: a correct translation that uses different synonyms can score poorly. That is why neural metrics such as COMET (Rei et al., 2020) have gained ground, comparing source, hypothesis and reference in a semantic space and correlating better with human judgment.

Real limits

For all the progress, machine translation is far from solved. Low-resource languages, without large bilingual corpora, still receive markedly poorer translations. Cultural nuance, humor, register and specialized domains —legal, medical— continue to demand human review. And a risk peculiar to neural models persists: hallucinations, fluent output disconnected from the source, most common precisely in low-resource directions. Machine translation in 2026 is an extraordinary tool, but not an infallible replacement for the human translator wherever precision is non-negotiable.