ALTO XML

ALTO XML enforces a strict separation between physical layout and logical content. The schema describes the precise spatial coordinates of text blocks, illustrations, and marginalia on a page without imposing any semantic interpretation. This allows a single ALTO file to serve as a neutral ground-truth layer for multiple downstream processing pipelines—legal document structure parsing, historical manuscript analysis, or newspaper digitization—each applying its own semantic overlay. The file records bounding box polygons for every recognized element, preserving the exact typographic geography of the original document.

ALTO decomposes a page into a hierarchy of nested elements: PrintSpace, TextBlock, TextLine, and String. Each String element represents a single recognized word and carries attributes for:

• CONTENT: The OCR-recognized text
• WC: Word confidence score (0.0–1.0)
• HPOS/VPOS: Horizontal and vertical position in pixels
• WIDTH/HEIGHT: Dimensions of the bounding box
• STYLEREFS: Pointer to font style definitions This granularity enables downstream token classification and named entity recognition systems to operate with precise spatial context.

ALTO captures detailed typographic metadata through TextStyle and ParagraphStyle elements referenced by text blocks. Attributes include:

• FONTFAMILY: Typeface name (e.g., Times New Roman)
• FONTTYPE: Classification (serif, sans-serif, monospace)
• FONTSIZE: Point size
• FONTSTYLE: Bold, italics, underline This metadata is essential for font-based heuristic parsing, where changes in font weight or size signal structural transitions—such as a bold 14pt heading followed by 11pt body text—enabling accurate header hierarchy extraction without semantic analysis.

ALTO provides explicit mechanisms for handling end-of-line hyphenation and logical reading order. The SUBS_TYPE attribute on a hyphenated word fragment indicates whether it is the first part (HypPart1) or second part (HypPart2) of a broken word, enabling accurate text reconstruction. The ID and IDREF attributes link related text blocks, while the optional ReadingOrder element defines the intended sequence of text regions—critical for multi-column legal documents where naive top-to-bottom extraction would scramble the content.

Beyond text, ALTO defines Illustration and GraphicalElement blocks for non-textual content. Each illustration carries:

• TYPE: Classification (map, photograph, drawing, chart)
• FILEID: Reference to an external image file
• POSITION: Bounding polygon coordinates This structured metadata enables zonal OCR pipelines to selectively ignore marginal illustrations, stamps, or seals while focusing processing on text-bearing regions. For legal documents, this is vital when extracting operative text from scanned contracts that contain logos, signature blocks, or notary seals.

ALTO is designed to function within the METS (Metadata Encoding and Transmission Standard) ecosystem. A METS file acts as a wrapper, referencing:

• Descriptive metadata (Dublin Core, MODS)
• Administrative metadata (provenance, rights)
• File sections pointing to TIFF/JPEG page images
• Structural map linking ALTO files to their corresponding images This integration allows legal digitization projects to maintain a complete preservation package where the raw scan, OCR output, and structural metadata remain permanently linked—essential for maintaining chain of custody and citation integrity in legal AI systems.

The algorithmic determination of the logical sequence in which text blocks, columns, and inset elements should be read. ALTO XML represents reading order through the ReadingOrder element, which groups OrderedGroup and UnorderedGroup references to block IDs. This is critical for legal documents where multi-column layouts, footnotes, and marginal citations create ambiguous visual paths. Graph-based approaches model text blocks as nodes with spatial relationship edges, then apply topological sorting to resolve the correct narrative flow.

A technique where optical character recognition is applied only to user-defined regions of a document, ignoring irrelevant areas like headers, footers, or marginalia. ALTO XML's Processing element records which zones were included or excluded during recognition. In legal document pipelines, zonal OCR enables:

• Selective extraction of operative provisions while skipping boilerplate
• Targeted re-OCR of low-confidence regions without reprocessing entire pages
• Privacy compliance by excluding personally identifiable information zones from recognition

The process of reconstructing logical document structure—paragraphs, headings, lists, reading order—from the unstructured stream of drawing commands in a PDF file. Unlike born-digital ALTO XML, PDFs store content as absolute-positioned glyphs with no inherent semantic hierarchy. Extraction pipelines must:

• Cluster spatially proximate glyphs into words and lines
• Detect font size/weight changes to infer heading boundaries
• Reconstruct table structures from ruling lines and cell alignments
• Output structured formats like ALTO XML for downstream legal NLP tasks

A token-level annotation standard used in sequence labeling where tokens are tagged as Beginning, Inside, or Outside of a named entity or structural chunk. When training models to generate ALTO XML from raw text, BIO tagging identifies:

• B-HEADING: First token of a section title
• I-HEADING: Continuation tokens of the same title
• O: Body text tokens This scheme enables token classification models (e.g., BERT-based architectures) to predict document structure boundaries that map directly to ALTO's TextBlock and String elements.