A text layer is often added automatically behind scanned documents, and this layer makes it possible to search the file. The convenience is real but its reliability is limited, because the software does not leave a character it cannot read blank; it fills it with a character that resembles it. This is one of the most silent sources of error in translation.
What the layer is
The text layer is a copy of the text placed behind the image and invisible to the eye. What appears on screen is still the scanned image; the layer exists only for searching and copying. That the two may differ mostly does not occur to anyone.
Silent production of errors
When the software cannot recognise a character, it replaces it with the one it finds most similar, and this change produces no warning. Because the image looks right, the error only comes to light when the text is copied. For this reason the layer is not a source to be trusted directly.
Similar characters
Zero and certain letters, and one and certain signs, are regularly confused in character recognition. This confusion has serious consequences in number fields in particular, because a single digit can make the document unusable. Numbers are read from the image, not from the layer.
The risk in number fields
An error in a text field can mostly be understood from the context, but in number fields there is no such chance of correction. A wrongly recognised digit leaves no trace showing that it is wrong. Numerical fields are therefore always verified by hand.
Handwritten text
Character recognition performs far worse on handwriting than on printed text, and the text it produces is mostly meaningless. Handwritten fields are not read from the layer at all. These fields are read by eye and, if they cannot be read, are recorded as such.
Table structures
In tables character recognition mostly fails to keep the arrangement of rows and columns, and the cells get mixed up. A scrambled table gives wrong information even if it consists of the right characters. Tables are read from the image.
Multi-column pages
On two-column pages the text layer may read the columns as one, and the sentences run into each other. The resulting text looks fluent but is meaningless. The column layout is verified by eye.
Language setting
Character recognition is set for a particular language, and when the wrong language is selected the special letters are corrupted. The letters specific to Turkish are the group most affected by this corruption. The setting is checked.
Text on seals and stamps
The writing on seal impressions is the hardest area for character recognition, and the text produced is generally unusable. These fields are read by hand. If they cannot be read, this is written.
Effect of a watermark
When a watermark lies over the text, character recognition may mix the two texts together. The result is words that exist nowhere. The layer is not used on watermarked pages.
Where the layer is useful
In long documents the text layer is genuinely handy for finding where a term occurs. This use is only a pointer, and the place found is again verified by eye. The tool is used for searching, not for carrying text over.
The copy-and-paste risk
When text copied from the layer is taken straight into the translation, the errors are carried along silently. This method is tempting because it looks fast, but it is the riskiest one. Every copied field is compared with the original.
Invisible characters
Copied text may carry formatting characters invisible to the eye, and these cause problems in later tools. The fields are cleaned by converting them to plain text. The cleaning is done at the start of the job.
Editable source
If the client has the editable source of the document, the character recognition problem disappears entirely. This file is requested at the start of the job because it both shortens the turnaround and lowers the risk of error. The request is always made.
The place of quality control
The quality of the character recognition output depends directly on the scan quality of the file. A poor scan cannot produce a good text layer. The two subjects are assessed together.
Tool-independent checking
Whichever tool produced the text layer, its output is verified independently. The assumption that the tool is reliable is exactly what lets errors of this kind slip through. Verification is always done against the image.
Responsibility
The translation rests on the original document itself; the text layer is an intermediate product and does not count as a basis. This distinction is indicated in the delivery letter where needed. The limit is explained at the start of the job.
Delivery check
All fields containing a number, a date or a sum are separately verified against the image. This check catches most of the errors that stem from character recognition. The step is not skipped.