Does the algorithm need a bilingual dictionary?

Its defining length-based method can operate without lexical translation information, although other alignment systems incorporate such evidence.

Can two sentences correspond to one?

Yes. The standard allowed link types include both two-to-one and one-to-two relationships.

Does it support arbitrary sentence reordering?

The usual dynamic-programming alignment follows the ordered sequences. Extensive rearrangement requires different handling rather than assuming the standard path can represent every correspondence.

AI Glossary / Natural language and speech / Intermediate

Gale - Church alignment algorithm

Definition

The Gale - Church alignment algorithm identifies corresponding sentences or small sentence groups in two versions of a text, typically written in different languages. It uses a statistical relationship between character lengths and dynamic programming to find a consistent sequence of alignments. It aligns existing translations; it does not produce new translations.

What it means

A translation does not always preserve sentence boundaries. One long sentence can become two shorter sentences, and material can occasionally be omitted or added. Consequently, pairing sentence number one with sentence number one throughout a document can fail.
Gale - Church uses the observation that longer passages often translate into longer passages, while shorter passages often translate into shorter ones. The relationship is approximate and can depend on the language pair and material.
The method is associated with William A. Gale and Kenneth W. Church and their research on bilingual corpora. Its importance is methodological: useful alignment evidence can come from length patterns even without a dictionary-based comparison of every word.

How it works

Start with two corresponding texts whose sentence boundaries have been identified. Reliable paragraph or other block boundaries can restrict the alignment problem to manageable corresponding regions.
Measure sentence lengths consistently. Score candidate links using a model of expected length differences and preferences for link types. The standard formulation permits one-to-one, one-to-two, two-to-one, two-to-two, and omission or insertion links.
Use dynamic programming to compare possible sequences of links. Rather than choosing each pair independently, the algorithm finds a best-scoring path through the two ordered sentence sequences under the allowed transitions.
Recover the links from that path and inspect uncertain results. The model assumes a largely preserved ordering; major rearrangement, unreliable sentence splitting, and missing sections can complicate the task.

Three practical examples

A bilingual-corpus project proposes sentence correspondences before reviewers approve data for translation research.
A lexicographer uses aligned passages to inspect how an expression is rendered in another language.
A teaching experiment compares simple position-by-position pairing with a method that allows one source sentence to align with two target sentences.

Think of it this way

Think of matching two ordered strings of differently sized beads, where one bead on one string may correspond to two on the other. Similar combined sizes offer clues, and the whole sequence constrains the choices. The analogy stops at meaning: equal-sized beads, like equal-length sentences, do not prove that two passages say the same thing.

Worked example

Suppose a tiny source block contains sentences with character lengths 40 and 60. Its target block contains three sentences with lengths 42, 28, and 35. Assume an illustrative expected target-to-source ratio of 1.05.
One candidate aligns the first source sentence with the first target sentence: 40 × 1.05 = 42, exactly matching that target length. It aligns the second source sentence with the final two target sentences together: 60 × 1.05 = 63, and 28 + 35 = 63.
This creates a one-to-one link followed by a one-to-two link. Every sentence belongs to a group and the order is preserved.
The arithmetic shows why this path is plausible under a length model. It is not a complete Gale - Church score: the real calculation also uses variability and link-type preferences, and other allowed paths must be considered.
Most importantly, the numbers contain no semantic evidence. Replacing a target sentence with unrelated text of the same length would leave this simplified length comparison unchanged.

Illustrative case study

In a fictional project, Maple Manuals compares two language editions of equipment instructions. A translator has split some long source sentences, so matching purely by sentence number causes later pairs to drift.
The team runs a length-based alignment experiment within checked corresponding sections. Reviewers inspect the proposed split and merge links and discover one section present in only one edition.
They record that mismatch before deciding whether the material belongs in the bilingual collection. This imagined case shows why alignment belongs within a data-review workflow; it claims no measured accuracy and does not assume that every edition is a faithful sentence-level translation.

Common misconceptions and mistakes

Sentence alignment is not word alignment. A correct sentence group does not specify which individual words correspond.
The method is not a translation-quality metric. Plausible lengths can accompany incorrect, unrelated, or incomplete wording.
The standard model does not assume every sentence has exactly one counterpart.
Default length parameters from a historical implementation are not universal facts about all writing systems or domains.
Dynamic programming finds an optimum for the defined scores and allowed paths. That optimum can still be wrong when the assumptions or preprocessing do not fit the documents.

Business use

Businesses with existing multilingual documentation can use sentence alignment as a preparation step for a reviewed bilingual resource. It can help organize candidate correspondences for translation-memory construction or terminology analysis.
Check document versions, sentence segmentation, and missing sections before interpreting alignment scores. Sample ordinary links as well as difficult split, merge, and omission cases. A low-cost automatic proposal is useful only if its error pattern is suitable for the intended downstream use.

Check your understanding

Two candidate target passages have the same character length but different meanings. Can the length information alone determine which is the correct translation?

Answer: No. Their length evidence is identical. Other context, lexical evidence, or review is needed to distinguish their semantic correspondence.

Related terms

Parallel corpus
Dynamic programming
Machine translation
Translation memory
sentence segmentation
sentence alignment
word alignment
bilingual lexicography

Learn next

draw a small alignment grid and mark the movement for one-to-one and one-to-two links. Then study a lexical alignment method to see how word-level evidence can complement length-based information.

Alignment method

Sources

Gale and Church, “A Program for Aligning Sentences in Bilingual Corpora,” original 1993 paper: https://aclanthology.org/J93-1004.pdf. NLTK, official Gale–Church implementation documentation: https://www.nltk.org/api/nltk.translate.gale_church.html. The character-length example above uses invented data and is not a full probabilistic scoring calculation. Independent source review: Gale–Church’s original length-model and ordered-alignment introduction and NLTK’s documented transition priors were read.

Previous: Gain ratio Next: Galerkin approximation