Can a word have several links?

Yes, if the representation permits them. Some modeling conventions are more restrictive than the underlying linguistic relationship.

Is an unaligned word necessarily a translation error?

No. A meaning or grammatical function can be conveyed implicitly or by a larger construction. The passage needs contextual review.

AI Glossary / Natural language and speech / Intermediate

Bitext word alignment

Definition

Definition: Bitext word alignment identifies corresponding words or tokens in two texts that express substantially the same content in different languages. Its output is a set of links between positions, rather than a new translation.

What it means

What it means: A bitext contains paired language versions. Document alignment finds matching documents, sentence alignment finds matching sentences, and word alignment works within those pairs. These are different levels of correspondence, and errors at an earlier level can undermine the later analysis.
The links describe how a particular passage was translated. They are not a universal bilingual dictionary: the appropriate counterpart of a word depends on its sense and surrounding expression. Some words correspond to several words, and others have no separately expressed counterpart. Word order can also change across languages. A useful alignment representation must make its treatment of these cases explicit.

How it works

How it works: A pipeline first prepares corresponding passages and chooses a tokenization convention. Statistical alignment methods learn associations and positional patterns from parallel examples. Other approaches can use contextual representations or additional linguistic information. The model scores candidate correspondences, and a decoding or selection procedure turns those scores into links.
Some classical models restrict each token on one side to a single counterpart, possibly an empty position. Combining alignments computed in both directions can produce a less restrictive result. The chosen combination rule affects which links survive. Reviewers then inspect correspondence errors separately from translation quality: a fluent translation can still receive poor links, and accurate links cannot repair an incorrect translation.

Three practical examples

Three practical examples:
1. A bilingual terminology project highlights candidate equivalent expressions in translated manuals. A reviewer checks the complete phrase before adding anything to the approved terminology database.
2. A translation editor selects a word in one language and sees the corresponding span in the other. Ambiguous or missing links are shown as uncertain rather than forcing a misleading highlight.
3. A research team transfers annotations between parallel sentences. It evaluates the transferred labels against independently annotated examples because correspondence alone does not guarantee that the same grammatical label applies in both languages.

Think of it this way

Think of it: Word alignment resembles drawing threads between corresponding pieces of two versions of a message. The threads may cross, branch or leave a piece unconnected. The analogy is limited because meaning is not always divisible into separate pieces: an idiom or grammatical construction may need to be interpreted as a whole.

Worked example

Worked example: Number the English phrase “the red house” as the₁ red₂ house₃. Number its Spanish counterpart “la casa roja” as la₁ casa₂ roja₃. Under a straightforward word-level reference, the links are (1,1), (2,3) and (3,2), with each pair giving an English position followed by a Spanish position.
Suppose a position-matching baseline instead proposes (1,1), (2,2) and (3,3). Only its first link matches the reference. Its link precision is therefore 1/3, and its link recall is also 1/3, because there are three proposed links and three reference links. Matching equal positions fails on the reordered adjective and noun.
A second system proposes all three correct links plus the incorrect link (2,2). Its precision is 3/4, or 75%, while its recall is 3/3, or 100%. This toy calculation assumes one unambiguous reference set. Real annotation schemes may distinguish certain and possible links, so their evaluation formulas must not be replaced indiscriminately with this simplified calculation.

Illustrative case study

Illustrative case study: Fictional Cedar Manuals aligns a small collection of translated maintenance guides. Reviewers notice that the interface highlights only one word of several approved technical expressions. They inspect the tokenization and alignment policy before blaming the translators.
The team adds phrase-aware review, records unresolved correspondences and tests the interface on held-out guide sections. It also checks whether sentence pairs actually match after revisions. This invented case illustrates how an alignment feature can support editorial work while keeping the translator responsible for equivalence. It does not assert that a particular algorithm improves translation quality or that every technical phrase has a one-to-one counterpart.

Common misconceptions and mistakes

Common misconceptions and mistakes: Sentence alignment does not supply word links automatically. Equal token positions need not correspond, and identical spelling is insufficient evidence of equivalent meaning. A missing link can reflect a legitimate translation difference rather than an omission. Conversely, assigning every word a link can conceal genuine missing content. Neural attention weights also require care: an internal weighting pattern is not automatically a validated word alignment suitable for editorial or evaluation use.

Business use

Business use: Before adopting alignment, define the action that the links will support: highlighting, terminology discovery or annotation transfer have different error costs. Preserve the original passages and token offsets so reviewers can inspect the evidence. Test language pairs and document types separately, including reordered phrases, names and repeated words. Measure useful correspondences rather than only the number of links produced. Changes in tokenization or translation revisions should trigger suitable rechecking, because a stored position can point to the wrong word after the underlying text changes.

Check your understanding

Check your understanding: A system links the second word in every source phrase to the second word in its translation. Why can it fail even when both phrases are correct?

Answer: Languages can express corresponding content in different orders, so position equality does not establish word correspondence.

Related terms

Parallel corpus
Statistical machine translation
Tokenization
sentence alignment
bilingual lexicon
phrase alignment
translation divergence
annotation projection

Learn next

align a short pair of translated passages manually. Mark cases that require multiple words or have no separate counterpart. Then compare two tokenization conventions and explain how the token-level links change while the intended translation relationship remains the same.

Tokenization

Sources

Sources: Och and Ney, A Systematic Comparison of Various Statistical Alignment Models, 2003: https://aclanthology.org/J03-1002.pdf . Jurafsky and Martin, Speech and Language Processing, chapter 13, Machine Translation: https://web.stanford.edu/~jurafsky/slp3/13.pdf . The house example, link calculations and Cedar Manuals case are original teaching material; the simple metric is not presented as the paper's ambiguity-aware alignment error rate.

Previous: Bitemporal modeling Next: Bitmap index