Chapter 10: Machine Translation
Master machine translation from parallel corpora and word alignment through sequence-to-sequence models with attention to modern transformer-based systems. Learn BLEU evaluation, multilingual transfer, and how to build translation pipelines.
Chapter Overview
Machine translation (MT) is the task of automatically converting text from one natural language to another. It is one of the oldest and most commercially impactful applications of NLP, powering services that handle billions of translation requests daily.
The field has progressed through three major paradigms. Rule-based MT (1950s--1990s) relied on linguistic rules and bilingual dictionaries. Statistical MT (1990s--2014) learned translation probabilities from aligned parallel corpora, using word alignment models (IBM Models 1--5) and phrase-based decoding. Neural MT (2014--present) replaced the entire pipeline with end-to-end neural networks, first using sequence-to-sequence models with attention, then transformers. The original Transformer architecture was itself proposed as a translation model.
This chapter covers the full MT pipeline: building and using parallel corpora, aligning words between languages, encoding and decoding with attention, training transformer-based translation models, evaluating translation quality with BLEU and its successors, and scaling to multilingual systems that handle dozens of languages simultaneously.
Chapter Roadmap
Click any topic to jump in
Parallel Corpora
The bilingual training data that powers all MT — how corpora are built, aligned, and filtered for quality.
Word Alignment
IBM Models and EM-based learning of word-to-word translation correspondences between languages.
From alignment to attention — two architecture paths
Seq2Seq + Attention
Encoder-decoder neural translation with Bahdanau and Luong attention mechanisms.
Transformer MT
Multi-head self-attention and parallel training that replaced recurrent translation models.
Evaluating quality and extending to many languages
BLEU Score
N-gram precision with brevity penalty — the standard automatic metric for translation quality.
Multilingual Models
mBERT, XLM-R, and zero-shot cross-lingual transfer across dozens of languages.
Sign up to unlock this chapter
This chapter is part of PixelBank Premium. Create a free account, then upgrade to read the full lesson — concepts, walkthroughs, and exercises.