Chapter 13
Machine Translation: Meaning, Search, and Evaluation
Chapter: Machine Translation, in Daniel Jurafsky and James H. Martin, Speech and Language Processing, 3rd edition draft, August 19, 2026.
Topics to learn
Page numbers refer to the linked chapter PDF.
- Why languages do not line up p. 2
- Word order p. 3
- Lexical splits and gaps p. 4
- What a language must express p. 5
- Encoder-decoder translation p. 6
- Tokenization p. 7
- Parallel training data p. 8
- Cross-attention to the source p. 9
- The information boundary between source and future target tokens p. 9
- Search p. 11
- Beam search p. 11
- Minimum Bayes risk decoding p. 14
- Why a greedy first token can miss a better complete translation p. 11
- Low-resource settings p. 16
- Data augmentation p. 16
- Multilingual models p. 17
- Who the system leaves out p. 17
- Evaluation, bias, and ethics pp. 18-21
- Human adequacy and fluency judgments p. 18
- Overlap metrics such as chrF and BLEU p. 18
- Embedding-based metrics p. 21
- Bias and other harms p. 21
FAQ
Is it leakage for the first target position to attend to the end of the source?
No, when the full source sentence is the input. The forbidden information is a future target token. Streaming translation has a different source boundary.
Does a high character-overlap score prove the translation kept a negation?
No. A short added or missing negation can reverse the meaning and leave most character n-grams intact. Overlap is not adequacy.
What failure is beam search meant to reduce?
A locally likely first token can lead to a worse complete translation. A beam keeps several prefixes so a slightly worse start can survive.
What is a typological divergence?
A systematic difference the translation has to handle, such as word order, a lexical split, or information one language must state and the other can omit.