Does NMT translate word by word?

It may generate tokens sequentially, but each decision can use wider source and target context.

Is a transformer the same as NMT?

No. A transformer is an architecture used for several tasks; translation is one application.

Can a model translate an ambiguous source sentence correctly?

It may choose a plausible interpretation, but missing context can leave multiple valid readings. Additional context or clarification can be necessary.

AI Glossary / Natural language and speech / Intermediate

Neural machine translation

Definition

Neural machine translation, or NMT, uses neural networks to produce a translation from one human language into another. The network learns patterns connecting source text with target-language expressions. It generates a translation using those learned relationships and the available context, rather than relying only on an explicitly written set of translation rules.

What it means

Translation requires more than replacing each source word with a dictionary equivalent. Languages differ in word order, agreement, idioms, and the information that their grammar makes explicit.
NMT models represent the input numerically and use that representation to guide the output. A translation can therefore rearrange or combine information across positions in a sentence.
Neural translation is a broad approach, not the name of one architecture. Recurrent encoder - decoder models were influential in its development, and transformers provide another important design. A general language model can also translate, but its ability must still be evaluated for the intended language pair and task.

How it works

A typical supervised training setup uses parallel text: source passages paired with corresponding translations. The texts are divided into tokens, often smaller than complete words, and the model learns to predict target tokens conditioned on the source and preceding target context.
In an encoder - decoder design, the encoder represents the source while the decoder constructs the target. Attention lets the decoder use relevant source information as it generates successive tokens. The original transformer implements this with attention-based components rather than recurrent processing.
At inference, a decoding procedure selects an output from the model's possible sequences. Greedy decoding chooses the locally highest-probability token; beam search retains several partial alternatives. Neither a high model score nor a wider search automatically guarantees a faithful translation.
The resulting text needs evaluation for meaning preservation, readability, terminology, omissions, and additions in the relevant setting.

Three practical examples

A support team translates incoming messages to help staff identify the customer's issue, retaining access to the original wording.
A localization team produces draft product descriptions that editors check for terminology and suitability for the intended audience.
A reader translates an unfamiliar-language article to understand its main content and checks important details against reliable language expertise.

Think of it this way

Think of a translator who has studied many paired examples and learned how expressions tend to correspond across languages. The person considers surrounding words rather than consulting a separate dictionary entry for every word.
The analogy is limited: a neural model does not acquire a person's lived understanding or professional accountability simply by processing examples. It can produce fluent wording while omitting a condition, changing a number, or choosing the wrong sense.

Worked example

Consider an original toy decoding calculation for a fixed source message. Two possible English outputs are “Package arrived” and “Parcel arrived.” Treat each word as one token and include an end token.
Suppose the first path has probabilities 0.6 for Package, 0.4 for arrived given Package, and 0.9 for ending. Its sequence probability is 0.6 × 0.4 × 0.9 = 0.216.
The second path has probabilities 0.4 for Parcel, 0.9 for arrived given Parcel, and 0.9 for ending. Its sequence probability is 0.4 × 0.9 × 0.9 = 0.324.
Choosing Package at the first step because 0.6 exceeds 0.4 can miss the higher-scoring complete alternative. This compares only the two stipulated paths; it is not a full search over a real model. Both candidates also require a meaning check against the source. Their scores are not measured probabilities of being correct translations.

Illustrative case study

In a fictional project, Harbor Tools uses NMT to draft equipment descriptions. Reviewers find that the sentences read smoothly, but one recurring product term is translated inconsistently and a conditional availability statement becomes unconditional.
The team builds a representative review set that includes these cases and compares revised workflows on terminology and preserved conditions. It retains source text beside each draft so reviewers can inspect the actual claim.
This invented case illustrates a quality-control decision. It makes no claim about a specific translation provider or a measured business saving.

Common misconceptions and mistakes

Do not equate fluency with faithfulness: a grammatical sentence can convey the wrong message.
Do not assume that neural translation always requires a separate encoder - decoder architecture; neural approaches can be organized differently.
Do not treat attention weights as a complete explanation of why a translation was selected.
Do not assume that more parallel data helps equally when the added data has the wrong domain or poor alignments.
Finally, avoid treating one benchmark score as a universal ranking across language pairs, document types, and error consequences.

Business use

Businesses can use NMT to support information access and translation production. The relevant question is whether the complete workflow delivers usable translations for the intended content, including any review effort.
Measure performance on representative material and inspect errors that matter to the organization: product identifiers, quantities, conditions, names, and consistent terminology. Keep a distinction between faster draft generation and faster completion of an acceptable translation. The latter includes corrections and unresolved ambiguities.

Check your understanding

A candidate has token probabilities 0.5, 0.8, and 0.5, including its end token. What is its sequence probability, and does that establish translation accuracy?

Answer: The product is 0.2. It is a model score for the specified sequence, not proof that the source meaning has been preserved.

Related terms

Machine translation
Parallel corpus
Attention
Transformer
Beam search
Post-editing
encoder–decoder model
translation evaluation

Learn next

compare greedy decoding with beam search using a small probability tree. Then inspect two fluent translations of one source passage and identify any differences in conditions, quantities, or implied meaning.

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, chapter 13, Machine Translation: https://web.stanford.edu/~jurafsky/slp3/13.pdf Vaswani and colleagues, Attention Is All You Need, original paper: https://arxiv.org/html/1706.03762v7 The numerical paths and fictional case are original teaching examples, not benchmark results. Independent review: SLP training/decoding and original transformer architecture passages consulted.

Previous: Neural-linear approximation Next: Neural modeling fields