How do you translate a PAGE XML OCR transcription without damaging the page geometry or confusing translated text with historical evidence? PAGE XML, short for Page Analysis and Ground Truth Elements, is used in document image analysis and transcription projects to represent page regions, lines, words, glyphs, reading order and recognized text. For libraries, archives and digital-humanities teams, multilingual PAGE XML localization is a practical data-integrity problem as much as a language problem.
Imagine a collection of handwritten letters digitized in high resolution. One team transcribes the original Italian handwriting, another corrects the line segmentation, and translators prepare English explanations for visitors. If the English text is simply written into the source transcription, the record no longer tells a researcher what the historical page said. If region coordinates are shifted to fit translated sentences, the annotations no longer point reliably to the pixels in the original image. Both failures can occur even when the English writing sounds excellent.
This guide explains PAGE XML pagecontent structure, the 2019-07-15 schema family, PcGts, Page, Coords, TextRegion, TextLine, Word, Glyph, TextEquiv and ReadingOrder in clear practical language. It provides translation decisions, OCR quality assurance, multilingual document examples and tests for avoiding damage to primary-source evidence. It is not a general guide to ordinary document translation, nor a replacement for ALTO XML-specific QA, METS packaging or IIIF search-result delivery; its subject is the relationship between page content, recognition data and the underlying image.
The essential model: image, structure, reading and translation
Picture four transparent sheets placed over the same archival page. The bottom sheet is the image: physical marks, dimensions and layout. The next sheet identifies regions, lines and reading order. The third is the source-language reading: machine recognition, manual correction or verified ground truth. The top sheet offers translations and explanations for visitors. PAGE XML is especially useful for the middle sheets, where text and structure are linked to pixels. Problems arise when the top sheet is accidentally written over the third.
What PAGE XML is actually for
A digitization project has images of newspapers, manuscripts and ledgers whose visible marks need machine-readable structure. Start by asking what the marks on the actual page can establish. A translation is a later claim about meaning; it should not silently replace the image-linked reading that supports that claim. PAGE XML can encode detected regions, text elements and reading sequences while linking them to the source image. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Decide which data describe the original page and which commentary will be offered in each language. Practical treatment should name the source field and its semantic owner before anyone edits it. If the field is used for recognition or geometry, editing it for readability may have unintended consequences. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Replacing source transcription with translated prose would erase the distinction between a witness and a later interpretation. Compare every edited text element with the actual pixels and keep any translation in an explicitly identified layer. A visual overlay is invaluable here because a file can remain readable XML while describing the wrong place on a page. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Versioned namespaces and schema identity
Two institutions exchange PAGE files but one uses a 2018 namespace and another expects 2019-07-15. A useful explanation should distinguish the physical source, a transcription of that source and an interpretation for the reader. Those layers may agree, but they are not interchangeable. The namespace and schema version determine the permitted structure, not merely the spelling of a software label. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Identify the version before parsing, validating or migrating; do not rewrite namespace URIs during localization. Design the data flow so recognition specialists and translators do not accidentally use the same field for different meanings. The boundary is an editorial decision supported by technical structure. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A global translation pass changes a schemaLocation URL and makes a valid document fail validation. Validate the source and result against the intended schema and document all deliberate version migrations separately. Do not accept a successful software import as proof that content survived the import faithfully. Inspect structure and meaning separately. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
PcGts as the document container
A scanned page arrives with metadata and recognition elements contained under PcGts. Thinking in page space is different from thinking in word-processing space. On the scanned page, the line’s physical location can remain fixed while sentences expand in translation. PcGts is an XML root element identifying the page-ground-truth structure and its dependent content. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Preserve element names, namespace declarations and the root’s structural obligations exactly. Explain the decision in ordinary language to future maintainers. A good local solution is one that another institution can recognize and evaluate without reproducing undocumented assumptions. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
An editor turns PcGts into a readable local-language label because it looks unfamiliar. Parse the resulting XML with a namespace-aware parser rather than judging the file by visual appearance. Combine automatic checks with a human reading of a representative original page. Both are needed for reliable digital evidence. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Page dimensions and image filenames
A repository links an OCR record to a 3508-by-4961-pixel source image. Historical evidence often includes inconvenient details. An apparent error, crossed-out word or unfamiliar abbreviation can be precisely what a later researcher needs to inspect. Page attributes such as imageFilename, imageWidth and imageHeight connect the annotation geometry to one raster. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Localize the human-facing file description, not the literal filename or numeric dimensions used by software. An editor can create a reader-friendly account without forcing it into the OCR structure. Side-by-side transcription and translation often communicate more honestly than an unexplained replacement. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A translated filename points to an image that does not exist, or dimensions are mistaken for millimetres. Load the linked image and verify that all annotated coordinates fall within the declared page bounds. Judge the release by whether it makes the original evidence easier to understand without making that evidence harder to recover. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
TextRegion and other region classes
A journal page has two columns, a photograph, a caption and an ornamental border. Different audiences will read the same object differently. The software needs reliable structure; a visitor needs readable explanation; a scholar may need the literal transcription. Region types distinguish textual evidence from images, graphic material and other page components. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Preserve region type semantics and describe unfamiliar content for visitors through separate editorial fields. When the source is difficult, record uncertainty instead of substituting a confident guess. A helpful note can explain a probable reading while preserving what the image actually shows. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Treating every detected region as a paragraph sends photograph noise into the translation workflow. Inspect a page overlay showing region type and position before sending text to a translator. An independent reader should be able to follow the link from a displayed translation back to the original image region that supports the claim. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Coords and polygonal boundaries
The source page contains a handwritten sentence that bends around a decorative illustration. Layout is part of evidence. A footnote, caption or margin entry may change the author, time or authority of the nearby words, even when all of them are translatable. Coords points record geometric boundaries as polygons in image coordinate space. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Do not replace polygon vertices with translated text positions or reinterpret comma-separated coordinate pairs as decimal values. Keep numerical and structural invariants in a controlled comparison. Translation review by itself is unlikely to detect a shifted polygon or a broken indexed region reference. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A language-aware number formatter silently changes the punctuation inside points and invalidates geometry. Render each polygon against the original raster and check that the visible region remains enclosed. Treat any disagreement between image, transcription and translation as a documented issue requiring the appropriate expert, rather than smoothing it away. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
TextLine segmentation
A line break divides two words because the source writer ran out of physical page space. Reliable multilingual publication requires a source that remains stable enough to compare versions. Changes to transcription and geometry deserve separate review from changes to display language. TextLine corresponds to recognized or manually segmented line units in the document image. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Keep line identity and source-space segmentation distinct from a translator’s preferred sentence wrapping. Make every change reversible in the sense that the original record is still retrievable. A reviewer ought to be able to reproduce the reading from the original scan and editorial notes. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Rewrapping English text inside the source TextLines falsely suggests that those lines existed on the original page. Display the transcription beside the raster and confirm each TextLine still matches its source baseline or polygon. Keep a short audit trail describing the specific page, region or record checked. This makes corrections reusable across institutions. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Word and Glyph hierarchy
A research project annotates handwritten words and sometimes individual characters to study recognition errors. The goal is not to preserve every original typo in the visitor-facing translation. It is to preserve the evidence that a typo, abbreviation or uncertain reading appeared in the source. Word and Glyph elements can encode more granular image-text alignment than complete lines. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Use them for recognition evidence and maintain a separate alignment plan for translations that expand or contract. Choose a test page that stresses the feature under discussion. The easiest clean paragraph usually cannot reveal failures involving complex scripts, interleaved regions or old typography. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
One translated word is inserted over three original words, creating an invented glyph correspondence. Count the source words and inspect token-to-image references before claiming word-level translation alignment. The test should tell reviewers not only whether a problem occurred but also which layer caused it: source capture, recognition, transcription, interpretation or export. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
TextEquiv and Unicode content
A line has one machine-recognized reading and another manually corrected reading. Start by asking what the marks on the actual page can establish. A translation is a later claim about meaning; it should not silently replace the image-linked reading that supports that claim. TextEquiv provides textual equivalents for image elements, often using Unicode text content. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Identify which equivalent is original OCR, correction or source-language transcription before adding any translated interpretation. Practical treatment should name the source field and its semantic owner before anyone edits it. If the field is used for recognition or geometry, editing it for readability may have unintended consequences. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
The corrected source is overwritten by English, and downstream users can no longer reproduce OCR error measurements. Compare each TextEquiv reading with documented origin and ensure the original transcription remains retrievable. A visual overlay is invaluable here because a file can remain readable XML while describing the wrong place on a page. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
OCR confidence versus translation confidence
A recognition engine assigns uncertainty to a faint surname in a damaged register. A useful explanation should distinguish the physical source, a transcription of that source and an interpretation for the reader. Those layers may agree, but they are not interchangeable. Recognition confidence concerns how reliably the image was read; translation uncertainty concerns how a reading was interpreted. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Record these different uncertainties separately rather than collapsing them into a single quality score. Design the data flow so recognition specialists and translators do not accidentally use the same field for different meanings. The boundary is an editorial decision supported by technical structure. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A fluent English guess is assigned a high OCR confidence despite the original glyph being undecipherable. Check both the visual reading and the translated claim and retain uncertainty markers where evidence is weak. Do not accept a successful software import as proof that content survived the import faithfully. Inspect structure and meaning separately. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
ReadingOrder and indexed references
A nineteenth-century newspaper uses a main column, a sidebar and a continuation box. Thinking in page space is different from thinking in word-processing space. On the scanned page, the line’s physical location can remain fixed while sentences expand in translation. ReadingOrder can represent a chosen sequence of regions independently of their visual coordinates. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Preserve ordered group and region references while localizing explanatory navigation text. Explain the decision in ordinary language to future maintainers. A good local solution is one that another institution can recognize and evaluate without reproducing undocumented assumptions. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A translator reorganizes elements into a smoother modern narrative and inadvertently changes the encoded historical reading order. Walk RegionRefIndexed references and verify that every referenced region exists and appears in the intended order. Combine automatic checks with a human reading of a representative original page. Both are needed for reliable digital evidence. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Nested regions and cross-column text
A page contains a title across two columns followed by separate articles and a footnote spanning the bottom. Historical evidence often includes inconvenient details. An apparent error, crossed-out word or unfamiliar abbreviation can be precisely what a later researcher needs to inspect. Hierarchical layout relationships may describe reading units that cannot be recovered from simple top-to-bottom sorting. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Retain the original grouping and document how translated presentation should navigate complex layouts. An editor can create a reader-friendly account without forcing it into the OCR structure. Side-by-side transcription and translation often communicate more honestly than an unexplained replacement. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A plain text export reads both columns line by line across the gutter and creates nonsensical sentences. Compare the reading order with a human reading of the source page and flag ambiguous continuations. Judge the release by whether it makes the original evidence easier to understand without making that evidence harder to recover. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Right-to-left and bidirectional scripts
A source manuscript combines Arabic words, European dates and marginal Latin quotations. Different audiences will read the same object differently. The software needs reliable structure; a visitor needs readable explanation; a scholar may need the literal transcription. Text direction and script handling affect reading and display but do not automatically change page geometry. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Use Unicode-aware editing and record language or script context in fields supported by the actual PAGE version and workflow. When the source is difficult, record uncertainty instead of substituting a confident guess. A helpful note can explain a probable reading while preserving what the image actually shows. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Visual reordering is mistaken for logical character order, producing reversed text in exported evidence. Test selection, copying and display with mixed-direction samples and compare logical strings, not only screenshots. An independent reader should be able to follow the link from a displayed translation back to the original image region that supports the claim. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Line break hyphenation
A printed word is divided across two physical lines at the edge of a narrow column. Layout is part of evidence. A footnote, caption or margin entry may change the author, time or authority of the nearby words, even when all of them are translatable. The source layout and a normalized reading may legitimately represent different segmentations. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Preserve the original line evidence and explain any reconstructed whole-word reading as an editorial operation. Keep numerical and structural invariants in a controlled comparison. Translation review by itself is unlikely to detect a shifted polygon or a broken indexed region reference. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A translator silently joins words and loses the evidence necessary to reproduce the page segmentation. Keep a trace from normalized words back to the physical line and the location of any source hyphen. Treat any disagreement between image, transcription and translation as a documented issue requiring the appropriate expert, rather than smoothing it away. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Marginalia and annotations
A handwritten notebook contains later reader notes in the margin alongside the author’s main text. Reliable multilingual publication requires a source that remains stable enough to compare versions. Changes to transcription and geometry deserve separate review from changes to display language. Source annotations may have different authors, dates, purposes and spatial relationships from the main writing. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Translate these voices separately and preserve any uncertain attribution rather than merging them into a single narrative. Make every change reversible in the sense that the original record is still retrievable. A reviewer ought to be able to reproduce the reading from the original scan and editorial notes. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A marginal correction is inserted into the main sentence without notice and changes what the original author appears to say. Show a side-by-side overlay and distinguish primary text, later writing and editorial translation. Keep a short audit trail describing the specific page, region or record checked. This makes corrections reusable across institutions. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Tables and ledgers
A shipping register uses columns for dates, names, goods, quantities and payments. The goal is not to preserve every original typo in the visitor-facing translation. It is to preserve the evidence that a typo, abbreviation or uncertain reading appeared in the source. Table layout conveys relationships that cannot be represented safely as a continuous translated paragraph. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Retain cell boundaries and numeric data while translating headings and explaining obsolete terminology. Choose a test page that stresses the feature under discussion. The easiest clean paragraph usually cannot reveal failures involving complex scripts, interleaved regions or old typography. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A row shifts under the wrong heading when labels become longer, making a payment belong to another trader. Test each column association and row identity against a reference image before exporting the translated display. The test should tell reviewers not only whether a problem occurred but also which layer caused it: source capture, recognition, transcription, interpretation or export. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Illustrations and non-text regions
A natural-history book includes drawings, scale bars and labels around the illustrations. Start by asking what the marks on the actual page can establish. A translation is a later claim about meaning; it should not silently replace the image-linked reading that supports that claim. Image regions and textual labels do different jobs even when they appear visually close. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Translate captions and alt descriptions while preserving measured scale and source positions. Practical treatment should name the source field and its semantic owner before anyone edits it. If the field is used for recognition or geometry, editing it for readability may have unintended consequences. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
OCR guesses on a decorative drawing are presented as real words in the translated edition. Inspect unusual regions manually and keep uncertain or non-text detections out of trusted transcription. A visual overlay is invaluable here because a file can remain readable XML while describing the wrong place on a page. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Historical spelling and abbreviations
A legal deed uses obsolete spelling, contractions and a clerk’s distinctive punctuation. A useful explanation should distinguish the physical source, a transcription of that source and an interpretation for the reader. Those layers may agree, but they are not interchangeable. The transcription records what is on the source; translation explains its meaning for another reader. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Retain diplomatic transcription and provide expansion or interpretation with an explicit editorial status. Design the data flow so recognition specialists and translators do not accidentally use the same field for different meanings. The boundary is an editorial decision supported by technical structure. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Modernizing the original text destroys evidence useful for linguists, historians and handwriting researchers. Ask whether every expanded abbreviation is grounded and whether the unexpanded source remains available. Do not accept a successful software import as proof that content survived the import faithfully. Inspect structure and meaning separately. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Named entities and uncertain names
An OCR system reads a family name inconsistently across related parish pages. Thinking in page space is different from thinking in word-processing space. On the scanned page, the line’s physical location can remain fixed while sentences expand in translation. A normalized authority name and the written string on the page are not the same object. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Link an identified person through an authority record without rewriting the transcribed surname. Explain the decision in ordinary language to future maintainers. A good local solution is one that another institution can recognize and evaluate without reproducing undocumented assumptions. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A plausible modern spelling is substituted into TextEquiv and presented as original writing. Preserve the literal form, record the authority match separately and state uncertainty where attribution is tentative. Combine automatic checks with a human reading of a representative original page. Both are needed for reliable digital evidence. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Image transformations and derivative sizes
An OCR file was generated against a high-resolution raster, but the website displays a reduced thumbnail. Historical evidence often includes inconvenient details. An apparent error, crossed-out word or unfamiliar abbreviation can be precisely what a later researcher needs to inspect. Coordinates are meaningful only relative to the image dimensions and transformations for which they were defined. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Translate text independently while preserving a documented transform for any new visualization scale. An editor can create a reader-friendly account without forcing it into the OCR structure. Side-by-side transcription and translation often communicate more honestly than an unexplained replacement. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Region coordinates are copied unchanged to a thumbnail and annotations float far from their targets. Compare a known control region at original and display scale and test the coordinate mapping quantitatively. Judge the release by whether it makes the original evidence easier to understand without making that evidence harder to recover. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Ground truth for OCR evaluation
Researchers use carefully corrected PAGE files to evaluate line recognition on a multilingual historical corpus. Different audiences will read the same object differently. The software needs reliable structure; a visitor needs readable explanation; a scholar may need the literal transcription. Ground truth should represent verified source evidence and the evaluation scope, not whichever text reads most naturally today. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Store translation as a parallel research product and keep the benchmark reference clean. When the source is difficult, record uncertainty instead of substituting a confident guess. A helpful note can explain a probable reading while preserving what the image actually shows. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A corrected English edition replaces source-language gold data and makes recognition accuracy comparisons meaningless. Run the expected evaluation against the original reference and check that all benchmark identifiers remain unchanged. An independent reader should be able to follow the link from a displayed translation back to the original image region that supports the claim. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Schema validation and semantic review
A file passes XML syntax checking but the lines appear over the wrong photograph. Layout is part of evidence. A footnote, caption or margin entry may change the author, time or authority of the nearby words, even when all of them are translatable. Schema validity protects structural constraints but cannot by itself guarantee semantic correspondence with pixels. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Combine schema checks, image overlays, reading-order inspection and linguistic review. Keep numerical and structural invariants in a controlled comparison. Translation review by itself is unlikely to detect a shifted polygon or a broken indexed region reference. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A team approves a well-formed file although every TextLine polygon was shifted after deskewing. Use representative manual visual audits together with automated out-of-bounds and broken-reference checks. Treat any disagreement between image, transcription and translation as a documented issue requiring the appropriate expert, rather than smoothing it away. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Interchange with ALTO and IIIF
A library wants to use PAGE annotations for OCR editing, ALTO in a delivery system and IIIF for viewing. Reliable multilingual publication requires a source that remains stable enough to compare versions. Changes to transcription and geometry deserve separate review from changes to display language. The three systems can cooperate but do not have identical data models or lossless field mappings. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Record explicit crosswalk decisions and retain original identifiers and evidence where mappings discard detail. Make every change reversible in the sense that the original record is still retrievable. A reviewer ought to be able to reproduce the reading from the original scan and editorial notes. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A conversion claims perfect fidelity after dropping glyph annotations and complex reading-order groups. Compare a sample page before and after conversion for text, region boundaries, sequence and provenance. Keep a short audit trail describing the specific page, region or record checked. This makes corrections reusable across institutions. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Provenance and editorial responsibility
One digitization project includes scans, raw OCR, corrected transcription, translation and explanatory annotations. The goal is not to preserve every original typo in the visitor-facing translation. It is to preserve the evidence that a typo, abbreviation or uncertain reading appeared in the source. Each layer was produced by different decisions and may have different dates, confidence and ownership. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Keep edition and process metadata sufficient to distinguish recognition output from later scholarly interpretation. Choose a test page that stresses the feature under discussion. The easiest clean paragraph usually cannot reveal failures involving complex scripts, interleaved regions or old typography. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
A polished translation is labelled as the manuscript’s original text with no trace of editing. Ask a new researcher to reconstruct the source and corrections without privileged background knowledge. The test should tell reviewers not only whether a problem occurred but also which layer caused it: source capture, recognition, transcription, interpretation or export. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
A resilient multilingual release
An archive plans to publish German, English and Chinese displays of the same fragile manuscript. Start by asking what the marks on the actual page can establish. A translation is a later claim about meaning; it should not silently replace the image-linked reading that supports that claim. The safest release preserves source PAGE records and treats localized presentation as linked derivatives. The question is therefore twofold: what must remain true about the source image, and which part of the reader’s explanation may change when the language changes? Answering both avoids the false choice between accurate data and accessible prose.
Pilot one complex page from each document type before expanding to the full collection. Practical treatment should name the source field and its semantic owner before anyone edits it. If the field is used for recognition or geometry, editing it for readability may have unintended consequences. The most practical habit is to keep a source reference beside each human-readable derivative. It gives a translator enough context to write naturally while allowing an editor to check that the wording reflects the historical or documentary evidence, not merely the current software display.
Bulk updates are launched after checking only simple single-column pages, so marginalia and tables break. Require successful overlay, text and reading-order checks in every supported language before launch. A visual overlay is invaluable here because a file can remain readable XML while describing the wrong place on a page. A review report should name the specific invariant and show how the result was verified. Saying that an XML file opened successfully is much weaker than showing that the correct line, field, boundary and textual reading were preserved.
Worked example: a multilingual newspaper page
Consider a page containing a title across two columns, two articles, a photographic caption and an advertisement. The PAGE record points to a source image with declared pixel dimensions. A TextRegion represents the headline, other regions represent the articles, and a reading-order group expresses which region belongs before the next. A translation into English could reasonably join a sentence that spans several physical lines; it must not pretend the original printer arranged the page in the same English line breaks. The correct outcome is an English reading experience connected to the original regions, not a rewritten record of the historical page.
Build a test set containing the page image, original PAGE file, corrected source transcription and the new translation. Render every polygon over the image. Walk the reading-order references in sequence, looking for a column transition or caption inserted at the wrong point. Then compare the literal source reading with the image. Finally, review the English against that approved reading. The stages are deliberately separate: the translator should not be asked to approve a damaged polygon, and the OCR engineer should not be asked to judge the nuance of a historical legal phrase.
Repeat the exercise with a right-to-left source and a table. The second sample exposes direction and Unicode handling; the third checks whether row and cell associations remain intact. If translation is displayed in a separate IIIF viewer, trace the identifiers and coordinates all the way into the viewer rather than assuming conversion is lossless. Preserve an audit note about each transformation and unresolved reading. This is the standard of evidence that makes a multilingual edition useful beyond its launch day.
The complete quality-assurance route
Before delivery, verify the schema namespace and parse the XML. Confirm that the filename and image dimensions match the declared raster. Check every region and text polygon is in bounds. Resolve every reading-order reference. Compare the source transcription with the pixels and record uncertain readings. Review named entities and abbreviations without silently modernizing them. Confirm that translation is traceably separate from source ground truth. Test complex pages, mixed scripts and tables. Finally, inspect the exported view that real readers will see, especially if PAGE has been mapped into a different delivery format.
A small controlled sample is more informative than a huge undifferentiated pass. Choose cases that contain the features most likely to break: marginalia, columns, crossed-out words, low-confidence handwriting, decorative initials, non-Latin text and images mixed with captions. Record the exact source page and the expected reading. When a mismatch appears, identify whether it came from segmentation, OCR, transcription, translation or conversion. Fix the responsible layer and rerun the same check. This discipline avoids accidental changes to source records while keeping the public text genuinely understandable.
Frequently asked questions
Is PAGE XML the same as an ordinary translated Word file?
No. PAGE XML records page-image structure and recognized text, including geometric and reading-order information. A translation should not be mistaken for the original transcription or mapped naively onto source pixels.
Can I replace the Unicode text inside TextEquiv with English?
Only if a clearly defined application and editorial policy explicitly calls for that transformation and keeps the original evidence accessible. For source-ground-truth workflows, overwriting the source reading is usually the wrong choice.
Why does reading order matter?
Multi-column pages, marginal notes and tables may not read correctly in simple coordinate order. ReadingOrder references give downstream applications a deliberate sequence to follow.
Does XML validation prove that the translation is correct?
No. It proves certain structural requirements but does not verify that a text line matches the source image or that a historical phrase was interpreted responsibly.
How can PAGE XML be used with IIIF?
PAGE can preserve OCR-side structure; a separate integration can transform supported regions and text into annotations or search results associated with IIIF Canvas resources. The crosswalk needs testing.
What should a translator preserve first?
Source identity, image linkage, region coordinates, line and word references, reading-order logic and a recoverable distinction between original transcription, corrections and translation.
Further reading and related lessons
Consult PRImA Research Lab PAGE-XML documentation and schemas, PRImA page-layout examples and the related eduKateSG article on METS 2 digital library packages. For source discovery and viewer presentation, see IIIF Content Search API 2.0 and IIIF Presentation API 3.0.