Translation Memory Risk: Bowker’s Warning About Blind Faith

Translation memory is one of the most useful inventions in translation technology. It reuses sentences that have already been approved, holds terminology consistent and makes it possible to deliver updates within a reasonable time. Talking about the risk of a tool this useful may seem odd.

Contents

Yet translation memory risk is a real and measured phenomenon. The risk arises not from what is in the memory but from the form the trust placed in it takes. A suggestion marked in green on the screen gives the impression of a sentence whose accuracy has been proven; in fact it is only a sentence that has been used before.

In this article we set out why the memory is so persuasive, what Lynne Bowker’s warning about blind faith says, how an error multiplies by way of the memory, and how context checking closes the risk.

Why is translation memory so persuasive?

In a CAT tool the memory suggestion arrives in front of the translator as a ready-made answer. The sentence has already been written, the grammar is sound, the terms are in place. Accepting it rather than assessing it is both faster and cognitively cheaper.

The interface strengthens this tendency. The match rate is shown as a percentage, and the phrase a hundred per cent carries a connotation of correctness. Yet that figure measures only the similarity of the source text, not the fitness of the target text for this document.

The economics push in the same direction. In many pricing models exact matches are counted as discounted or are not charged for at all. A translator who reads an exact match carefully is not paid for doing so; a translator who moves past it quickly is noticed by nobody.

To this is added the pressure of volume. Industry research puts the average daily output of translators at 2,684 words, against a common industry benchmark of 2,500 words (DePalma 2012). At that pace, questioning every exact match afresh is not realistic; the ones to be questioned have to be selected.

Bowker’s warning about blind faith

In a pilot study published in 2005, Lynne Bowker gave this tendency a name: blind faith in the translation memory (2005 data). Bowker observed that the entries held in a memory were largely accepted by translators without question, and that even faulty entries were carried over into their translations.

The design of the study was simple. Flawed translations deliberately planted in the memory survived largely intact in the texts the participants produced. The same flaws did not appear when the work was done without a memory. What produced the error, in other words, was not the translator’s knowledge but the feeling of confidence the tool conferred.

Bowker’s conclusion was not to reject the technology. The title of the study asked about the trade-off between productivity and quality; her answer was that the trade-off can be managed. The means of managing it are training and process, not the tool itself.

The finding matters more now than it did then. Once machine translation suggestions were added alongside the memory, the quantity of ready-made text put in front of the translator multiplied. The risk of blind faith works in exactly the same way whatever the source of the suggestion.

How translation memory risk grows: the multiplying of error

The most important property of a memory is that it is cumulative. Every segment that is approved is written into the memory and comes back as a suggestion in later projects. For correct translations this is a great gain. For faulty ones it is a silent mechanism of multiplication.

Once the wrong rendering of an institution name has entered the memory, that rendering is offered as an exact match in every document that follows. After the third document nobody questions it any longer, because it is consistent. Consistency begins to be taken for accuracy.

The second route to multiplication is the merging of memories. When entries from different clients, different years and different quality levels are gathered into a single file, the weakest entry acquires the same authority as the strongest. A memory of uncertain provenance is a memory that cannot be audited.

The third route is alignment. Turning old translations into a memory by matching them with their source texts is a quick method, but badly aligned lines enter the memory as mismatched pairs. Unless those entries are weeded out afterwards, they stay in circulation for years.

Context drift: the right sentence in the wrong place

The subtlest form of the risk is not the faulty entry but the displaced one. The sentence is entirely correct in its original context; in the new document it produces the wrong meaning. The table below summarises how the risk is distributed across the types of match.

Type of match	Confidence conveyed by the tool	Actual risk	Check required
Context match	Very high	Low	Sample reading
Exact match	High	Medium	Full check on headings and short phrases
High fuzzy match	Medium	High	Reading the whole of the difference
Low fuzzy match	Low	Medium	Weighing up writing from scratch
Machine translation suggestion	Visually identical	High	Separate flagging and full correction

The row that stands out in the table is high fuzzy matches. When the difference between the source sentence and the suggestion is small, the translator can miss it; yet the single word that has changed is very often a critical element such as a negation, a date or an amount.

Short segments carry a risk of their own as well. A one-word heading can mean something else elsewhere in the document. In serious processes, headings and table cells are therefore gone over regardless of the match rate.

Maintaining the memory: whose responsibility is it?

A translation memory is not an asset that improves of its own accord once it has been set up. Left unmaintained, it decays over the years: institution names change, statutory terms are updated, client preferences shift. The memory, meanwhile, carries on giving suggestions exactly as it did on the first day.

The first step in maintenance is separation. Keeping separate memories by client, by field and by year lowers the match rate somewhat but raises auditability markedly. A memory in which nobody knows where a given entry came from cannot be put right.

The second step is writing corrections back. If an error found at final review is corrected only in the delivery file, the same error will be suggested again in the next project. Feeding the correction into the memory is the one step that makes the quality improvement permanent.

The third step is periodic cleaning. When an institution name, a job title or a statutory term changes, a bulk search and replace is run over the memory. That single operation heads off dozens of errors that would otherwise surface in the next project.

The step that breaks blind faith: context checking

The practice that closes the risk is what we call context checking. The aim is not to retranslate every segment; it is to check, in a targeted way, whether the suggestions coming out of the memory are also right in this document. The checking is selective, because resources are finite.

In practice there are three priority areas. The first is numerical and legal elements: dates, amounts, article numbers, expressions of negation. The second is proper names: institutions, job titles, document names. The third is short segments: headings, table cells, button and menu text.

These three areas make up a small share of the translation volume but give rise to the bulk of the objections raised after delivery. Targeted checking closes most of the risk without incurring the cost of a full check.

The fourth step, and the one most often skipped, is making the origin of the suggestion visible. If suggestions coming from the memory and suggestions coming from machine translation are flagged separately, the translator directs their attention to the right place. A set-up that does not draw this distinction manufactures blind faith at the level of the tool.

Context checking does not cost as much as people assume. When the fields to be checked are defined in advance, the extra time stays a small share of the total translation time. Set against that, the risk it closes runs to far more expensive outcomes: correction after delivery, reprinting, or delay in an official procedure.

The Türkiye context

The first factor that enlarges memory risk in Turkish is how often institution names and job titles change. Ministry and agency names change with reorganisations; memory entries carrying the old name go on being offered as exact matches for years. In an official document, an institution name that has lost its currency can lead to the application being refused.

The second factor is the structure of the language. Turkish is agglutinative, and the suffix a word takes changes with its function in the sentence. A noun phrase that arrives from the memory in isolation can settle into a new sentence with the wrong suffix agreement. A phrase that stays the same in English can take five different forms in Turkish.

The third is that Turkish counts as a low-resource language. Finding a ready-made, vetted and domain-specific memory and termbase for Turkish is far harder than it is for pairs such as English-German. In the Türkiye market the quality of a memory therefore depends largely on an agency’s own accumulated work and on how regularly it cleans that work up.

What does this mean for you?

For the buyer of translation the inference here is this: the existence of a memory is not a quality guarantee, it is a means of production. What determines quality is how the memory is maintained and how the suggestions are checked. You can keep a check on this with a few concrete steps.

1. When your organisation name, your job titles or your product names change, notify your translation provider in writing; the memory should be updated accordingly.
2. Have it confirmed in the quotation that dates, amounts and article numbers are checked regardless of the match rate.
3. Have it written into the contract that the memory belongs to you and will be handed over on request.
4. Ask whether the corrections you approve are fed into the memory; a correction that is not fed in comes back in the next project.
5. If machine translation is used, find out that it is flagged separately and goes through full correction.

To see how your memory is managed and which checks fall within scope you can look at our translation services page, and to see how differences in scope are reflected in price, at our pricing page.

Frequently asked questions

Where exactly does translation memory risk arise?

Not in what the memory holds but in the trust placed in it. When the tool marks a sentence as an exact match, the translator tends not to weigh it up again. Even if the sentence is out of date, wrong or unsuited to this context, it looks approved.

Who introduced the notion of blind faith?

Lynne Bowker used it in a pilot study published in 2005 to describe unquestioning trust in the translation memory. The study showed that a faulty entry in the memory was carried over into students’ translations to a considerable degree.

Is an exact match always safe?

No. An exact match tells you only that the source sentence is word for word the same; it does not tell you that the sentence means the same thing in this document. Headings, table cells and short phrases may have to be translated differently according to context.

How can I check memory quality as a client?

Ask when the memory was built and which projects it was built from, and whether it has been updated after changes to institution names or job titles. State in the contract that your memory file can be handed over to you.

Sources

• Bowker, L. (2005). Productivity vs quality? A pilot study on the impact of translation memory systems. Localisation Focus, 4(1), pp. 13–20.
• Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation: Insights from a Case Study. London: Routledge Focus, pp. 20–34.
• DePalma, D. A. (2012). Translation Output Metrics. Common Sense Advisory industry research.
• Mossop, B. (2006). Has computerization changed translation? Meta, 51(4), pp. 787–793.