Translation memory (TM) is the tool mentioned most often in corporate translation and the one most often misunderstood. It is frequently confused with machine translation; yet a memory does not produce a new text, it only remembers sentences that a human has translated before. This distinction is decisive for both quality and confidentiality.
Definition
A memory is a database that stores pairs of source and target sentences. Each time a translation is completed, these pairs are added to the memory. In this way past jobs turn into an accumulated asset.
How it works
When a new text arrives, the memory compares every sentence with its own records and suggests the similar ones. A suggestion is a candidate, not an obligation. The translator makes the decision.
Confusion with machine translation
A memory does not produce new text; it only brings up the rendering that a human wrote earlier. Machine translation, on the other hand, produces the text on the spot. The two tools are entirely different things, and they also lead to different consequences for confidentiality.
How it differs from a termbase
A memory works at sentence level, a termbase at word level. The two complement each other but neither replaces the other. On a job the two are generally used together.
How it differs from a dictionary
A dictionary offers general knowledge of the language, whereas a memory carries the past decisions specific to that client. A rendering in the memory is a preference, not a universal truth. For this reason it is assessed together with its context.
Keeping it per client
Because every organisation has its own language, the memory is separated per client. A merged pool mixes up term preferences. The separation is built into the structure.
Match types
A memory produces exact, fuzzy and context matches, and each carries a different level of confidence. Fuzzy matches require separate review. The types are assessed separately.
Measuring the gain
The match rate coming from the memory can be measured, and it reduces the volume of work. This gain is shared with the client transparently. Hiding it costs trust.
Its value in recurring jobs
The gain is highest in periodic reports and template documents. Because the structure stays fixed, the match rate rises. The value increases over time.
How a wrong record multiplies
If a wrong rendering enters the memory, it is carried into all later jobs. The error does not stay in a single document. This is why maintenance is needed.
The need for maintenance
Old and faulty records are cleaned out regularly. A memory without maintenance loses its reliability over time. The operation is carried out to a plan.
Context blindness
A memory recognises the sentence but does not know the section it sits in. The same sentence may need a different rendering in a different section. This limit is known.
Sentences containing numbers
In sentences where the same pattern is repeated with different numbers, the memory may carry over the old number. This error is seen in tables in particular. Numbers are verified separately.
The confidentiality dimension
A memory stores the client's texts and therefore falls within the scope of confidentiality. Access is restricted and the memory is not shared. The rule is defined in the contract.
The question of ownership
Who the memory belongs to can be set out in the contract. If the client asks for it, the file is handed over. The decision is made clear at the outset.
File format
Memories can be exported in a portable format and moved between different tools. This reduces dependence on a tool. The format is standard.
Backup
A memory is backed up regularly, because losing it wipes out years of accumulated work. The backups are subject to the same access restriction. The routine is defined.
Common mistake
The most common mistake is to accept the suggestions coming from the memory automatically. That is to take the tool for a decision-making authority. The memory offers a suggestion, the translator decides.