Translation Segmentation: How Is a Text Split into Pieces?

Everyone who opens a translation tool for the first time sees the same thing: the text has turned into a two-column table. Source sentences on the left, empty boxes on the right. The process behind that display is called translation segmentation.

Contents

Segmentation is the operation of dividing a text into units that a translation memory can work with. It is a technical necessity: the memory matches at segment level, so without the division no match rate can be calculated at all.

But this technical choice has two significant consequences: it sets both the repetition discount on your invoice and the way the text is read. This article explains both.

What is translation segmentation?

When a translation tool opens a document, it first extracts the text and then breaks that text into pieces. The default unit of division is the sentence; headings, table cells, list items and image captions also count as separate segments.

The break point is generally set by punctuation: the full stop, the question mark, the exclamation mark, the colon. Paragraph ends and formatting boundaries produce breaks as well.

Once the segments have been created, each is looked up in the memory and a match rate is calculated. Where a segment begins and where it ends therefore determines directly whether that segment will be found in the memory.

This is why segmentation, an apparently technical detail, is the foundation of the entire process. A badly divided text produces both a low match rate and an unusable memory.

It is worth asking, too, why the sentence was chosen as the segment unit. A paragraph is too large a unit; a single word changed spoils the match for a whole block. A word is too small; it carries no context. The sentence is a reasonable balance between the likelihood of recurrence and wholeness of meaning. The balance is not perfect, but it has worked in practice for forty years.

How are segmentation rules determined?

Every tool arrives with a set of rules. These rules define which characters produce a break and in which cases a break is suppressed. The rules vary by language and can be edited by the user.

A common format has also been developed so that rule sets can travel between tools: SRX (Segmentation Rules eXchange). Rules defined in one project can therefore produce the same result in another environment.

The most important component of the rules is the exception list. The full stop in abbreviations such as “Dr.”, “vb.”, “Sn.” and “No.” is not a sentence end; these items are added to the list so that a wrong break is prevented. The richer the list, the better the quality of the segmentation.

In practice, tuning the rules to the individual project is often skipped. Yet article numbers in legal texts, or units of measurement in technical documents, are broken wrongly on a regular basis under the default settings.

Once the rules have been changed, a retrospective effect appears as well. When the same document is divided again under different rules, the records in the old memory may no longer line up with the new segments and the match rate drops. Rule changes should therefore be made at the start of a new family of documents, not in the middle of a project.

Where is the segment boundary drawn wrongly?

Segmentation errors recur in particular patterns. The table below summarises the most common ones and their effects.

Situation	What happens?	Result
Full stop in an abbreviation	The sentence is cut in two in the middle	Half sentences pile up in the memory
Decimal number	The number is separated into two segments	The number check throws an error
Article number	The number becomes a segment of its own	Context is lost, translation gets harder
Manually inserted line break	A single sentence is split into several pieces	The match rate falls
Table cells	A sentence is scattered across cells	The meaning is not visible inside the segment
Text embedded in an image	No segment is produced at all	The text is left out of the translation

Most of these errors arise from the way the source document was prepared. They occur not because the software is faulty but because the document is not as regular as the software expects it to be. The good news is this: most of them can be prevented with a few minutes of correction in the source file.

The last row of the table deserves separate emphasis. Because text embedded in images produces no segment at all, it does not show up in the analysis report either; the document is assumed to have been translated, while the labels in the diagrams have stayed in the source language. This is one of the gaps that costs most when it is noticed after delivery, and all it takes to prevent it is to list those texts in a separate file from the outset.

What is segmentation doing to the translation?

The technical benefit of segmentation is beyond dispute. Its effect on the text, however, has been debated for years, and that debate concerns the buying side too.

The first effect is that sentence boundaries become fixed. A good translator can merge two sentences of the source language into a single sentence in the target language, or split a long sentence in two. In a segmented environment that move becomes harder. Ehrensberger-Dow and Massey (2014) indeed report that sentence-merging behaviour drops by roughly 50% when translators work with a tool.

The second effect is a loss of wholeness. Mossop (2006) calls this “collage translation”: every segment is correct taken one by one, but nobody has read the text as a whole. Connectives weaken, references break, the rhythm of the paragraph is lost.

These two effects are mostly unimportant in technical documents; consistency comes before fluency there in any case. In promotional copy, in corporate communications and in content meant for publication, though, the outcome is noticeable. The answer is not to abandon the tool but to add a separate final review across the whole text to the workflow.

The third effect, and the one discussed least, is on the source text itself. As segmented production spreads, it pushes organisations towards shorter and more standardised sentences, because texts written that way yield higher matches. In technical documentation that is usually a good outcome. Carrying the same discipline into corporate communications, however, can flatten the language. It is better for the text to determine the tool than for the tool to determine the text.

How does segmentation affect the invoice?

Because match rates are calculated at segment level, every deviation at a segment boundary changes the analysis result directly. The same document can yield a different repetition rate under different segmentation rules.

A concrete example: if the paragraphs in the source document have been broken up with manually inserted line breaks, the full sentences in the memory can no longer be found. A text translated last year may yield almost no matches this year. The word count has not changed, but your invoice has gone up.

The reverse holds as well. Organisations that keep their source documents in the same layout from version to version obtain high context matches through that discipline alone. According to industry data, only around 59% of translation work is entirely new (DePalma 2012); how much of that potential you can use is largely set by the layout of your documents.

So if you see a large difference in repetition rates between quotations, ask first how the document was processed. Two agencies opening the same file with different settings can produce different analyses; that does not mean one of them is being more generous. To make the comparison fair, you need to ask for the analyses to be run on the same source file and the same memory.

Preparing the source text for segmentation

The steps below are the highest-return corrections that can be made in a source document before it is sent for translation. None of them requires special software.

1. Send the document in an editable format; look for the source file rather than a scanned PDF.
2. Clear out manually inserted line breaks inside sentences; handle alignment through the formatting settings.
3. Place tables as real tables rather than as images.
4. List the texts embedded in images in a separate file.
5. Write abbreviations consistently throughout the document and give your supplier the list.
6. Try to keep the general layout of the document intact between versions.

The total cost of these six items is usually under half a day, and the effect repeats with every new version. Document layout is the least discussed and most easily improved component of a translation budget.

The Türkiye context

Turkish carries two particular difficulties where segmentation is concerned. The first is sentence length: because subordinate clauses are attached to the main clause with suffixes in Turkish, a single sentence can become heavily layered. The likelihood of such sentences recurring word for word in the memory is lower than it is for short ones.

The second is the use of abbreviations and titles. Forms such as “Sn.”, “Av.”, “Dr.”, “vb.” and “T.C.” occur densely in official correspondence and, when they are not added to the exception list, produce wrong breaks on a regular basis. Using an abbreviation list prepared for Turkish brings a visible improvement on its own.

The third is the origin of the document. A significant share of the official paperwork sent for translation in Türkiye is scanned PDF; notarised documents, diplomas and registry records are often printouts with no source file behind them. An OCR stage is unavoidable with these documents, and the line breaks in the OCR output are the leading factor that disrupts segmentation. Correct recognition of Turkish characters calls for a separate checking step as well.

What does this mean for you?

Segmentation looks like a technical stage that falls under your supplier’s responsibility. In reality, what largely determines its outcome is the file you send.

In practical terms: sending the same content as a tidy Word file rather than as a scanned PDF changes the price, the delivery time and the quality alike. The difference is usually larger than people assume, and it is entirely within your control.

Ask for a final review across the whole text for anything that is going to be published, too. The disjointedness that segmented production creates can only be repaired at that stage, and it should be priced as a separate line of work. You can find detailed advice on document preparation in our knowledge centre, and request a quotation for the files you have to hand.

Frequently asked questions

What is segmentation and why is it done?

Segmentation is the operation of dividing a text into small units that a translation tool can work with. The default unit is generally the sentence. Without the division a translation memory cannot work at all, because the memory matches at segment level.

What happens if segment boundaries are drawn wrongly?

Abbreviations, decimal numbers and article numbers can produce wrong break points. The records in the memory are then corrupted, the match rate drops, and the translator is left working with meaningless fragments.

Does segmentation lower translation quality?

It can. Working sentence by sentence can lead to the text as a whole being overlooked. Research has shown that translators’ sentence-merging behaviour drops markedly when they work with a tool. A separate final review across the whole text is needed for that reason.

How should I prepare my source text to get a better result?

Send the text in an editable format, do not use line breaks inside sentences, do not place tables as images, and write abbreviations consistently. These four items raise the quality of segmentation visibly.

Sources

• Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation. Routledge Focus, Chapter 4.
• Ehrensberger-Dow, M. & Massey, G. (2014). “Cognitive ergonomic issues in professional translation”. The Development of Translation Competence. Cambridge Scholars.
• Mossop, B. (2006). “Has Computerization Changed Translation?”. Meta, 51(4), pp. 787–805.
• DePalma, D. A. (2012). Translation Productivity Benchmarks. Common Sense Advisory.