MORAIDICTIONNAIRE IA
NLP

Machine Translation

Automatically converting text from one language to another, from rules to neural networks.

Picture a tireless interpreter able to switch instantly between thousands of languages: that is the promise of machine translation (MT), the branch of natural language processing that converts text from a source language into a target language without human intervention.

From rules to neural networks

MT has gone through three great eras. The earliest systems (RBMT) relied on dictionaries and hand-written grammatical rules. Then came statistical models (SMT), which learned correspondences from huge aligned bilingual corpora. Since 2014, neural machine translation (NMT) has dominated: a network encodes the whole sentence into meaning vectors, then decodes it word by word.

Approach Foundation Strength
RBMT Linguistic rules Control, transparency
SMT Corpus probabilities Robustness
NMT Neural networks Fluency, context

The role of attention

The Transformer revolution (2017) rests on attention, which weighs the importance of each source word when generating each target word:

$$\text{Attention}(Q,K,V) = \text{softmax}!\left(\frac{QK^\top}{\sqrt{d_k}}\right)V$$

This mechanism captures long-distance dependencies and resolves ambiguities (does "bank" mean a riverside or a financial institution?) using the global context.

Measuring quality

MT is often evaluated with the BLEU score, which compares the n-grams produced against one or more reference translations. Yet such metrics remain imperfect: they struggle to judge naturalness, tone, or cultural fidelity, where humans still hold the edge.

Machine translation does not replace the translator: it multiplies their reach, while reminding us that meaning and nuance stay deeply human.

Explore the full AI dictionary →