COURSE 15L1100% FREE
Verified 2026-08-14

Machine translation basics

**Machine translation (MT)** maps text from one language to another. Modern MT is largely neural encoder–decoder / Transformer seq2seq. SLP3 includes a dedicated MT chapter; BART-style denoising pretraining also transfers to translation ...

What it is

Machine translation (MT) maps text from one language to another. Modern MT is largely neural encoder–decoder / Transformer seq2seq. SLP3 includes a dedicated MT chapter; BART-style denoising pretraining also transfers to translation settings (Lewis et al., ACL 2020).

<!-- IMAGE: source sentence → encoder-decoder → target sentence -->

HIGH PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Classic encoder–decoder MT: source sentence tokens into Encoder stack; context vectors; Decoder generating target tokens one-by-one with attention arrows from decoder step to source positions. Labels: 'source language', 'target language', 'attention'. Toy pair EN→ES 'the cat → el gato'.

Educational Focus: Encoder-decoder+attention is the Technical core and still teaches General learners the shape of MT.

Why it matters

Global business, support docs, and accessibility depend on MT. Errors in legal/medical domains can be costly. Fluent output can still be wrong—fluency ≠ fidelity.

How it works (plain)

  1. Encode source tokens into hidden states.
  2. Decode target tokens (autoregressive).
  3. Optionally constrain terminology with glossaries.
  4. Evaluate with automatic metrics and bilingual human review for critical content.
  5. For speech inputs, see Course 13 speech translation (cascades vs end-to-end).

CS224N treats Transformers and LLMs as core skills—the same stack powers modern MT systems students fine-tune or evaluate.

Everyday example

Translating a help center article, then having a bilingual human spot-check refund and safety pages.

Try it

Translate a paragraph round-trip (A→B→A). Note what drifted: names, negation, units, formality.

Myths

⚠️ Myth: Perfect fluency means perfect fidelity.
✓ Reality: Fluent mistranslation is dangerous.
⚠️ Myth: One BLEU point always equals better UX.
✓ Reality: Metrics are proxies—humans decide for high stakes.
⚠️ Myth: Multilingual LLMs remove the need for MT evaluation.
✓ Reality: You still need test sets per language pair and domain.

Sources