hOCR Format

hOCR encodes precise spatial information for every recognized element using the bbox property within the title attribute of span tags. This captures the x0 y0 x1 y1 coordinates of each word's bounding box, allowing downstream applications to reconstruct the original page layout without referencing the source image. This is critical for zonal OCR and optical layout analysis tasks where the physical position of a clause on a page carries legal significance.

A defining feature of hOCR is the x_wconf attribute, which exposes the OCR engine's confidence level for each recognized word as a percentage (0-100). This granular data enables:

• Flagging uncertain text for human review in high-stakes legal workflows
• Selective reprocessing of low-confidence regions with alternative OCR engines
• Weighted downstream analysis, where predictions from token classification for boundaries or named entity recognition can be calibrated based on input quality

hOCR leverages standard HTML elements to represent document hierarchy. div tags denote pages, p tags represent paragraphs, and span tags encapsulate individual lines and words. This design allows any standard Document Object Model (DOM) parser to navigate the output, eliminating the need for proprietary parsing libraries. The class attribute carries semantic meaning, with values like ocr_page, ocr_par, ocr_line, and ocrx_word defining the structural role of each element.

hOCR is the native output format for Tesseract OCR, one of the most widely deployed open-source recognition engines, ensuring broad compatibility. The format serves as a common interchange layer between the recognition phase and structural analysis tools like LayoutLM, which can consume the text and bounding box data jointly. This interoperability makes hOCR a foundational component in pipelines performing PDF structural extraction and header hierarchy extraction.

While hOCR excels at encoding text content and spatial layout, it has defined boundaries:

• Capabilities: Captures word coordinates, confidence scores, font attributes, and basic reading order
• Limitations: Does not natively represent complex semantic structures like tables, footnotes, or cross-references. For full logical structure, hOCR is often paired with complementary standards like ALTO XML or used as an intermediate step before transformation into Legal XML Schema formats.

In legal AI, hOCR acts as the critical bridge between scanned document images and machine-readable text. The format's ability to preserve exact spatial coordinates is essential for:

• Bates number extraction from fixed stamp locations
• Zonal OCR targeting specific regions like signature blocks
• Reading order detection in multi-column pleadings
• Providing the layout features required by multimodal models for section boundary detection

An open XML Schema maintained by the Library of Congress that describes layout and OCR information for text blocks, illustrations, and page elements. Unlike HOCR's HTML-based approach, ALTO uses a dedicated XML vocabulary with explicit measurement coordinates and block-level confidence scores. It is widely used in large-scale digitization projects and often serves as the interchange format between OCR engines and downstream text correction tools.

The computational process of segmenting a document image into regions of interest before text recognition occurs. This step identifies text columns, images, tables, and marginalia. HOCR encodes the output of this analysis as bounding boxes with class attributes (e.g., ocr_par, ocr_carea), making the spatial relationships between elements machine-readable and enabling downstream structural reasoning.

The algorithmic determination of the logical sequence in which text blocks should be read in complex layouts. HOCR supports this through the reading_order property on carea elements. This is critical for multi-column legal documents where a naive top-to-bottom, left-to-right extraction would scramble the intended narrative flow of a contract or judicial opinion.

The process of reconstructing logical document structure from the unstructured stream of drawing commands in a PDF. HOCR is often the intermediate representation generated after rendering PDF pages to images and applying OCR. The resulting HOCR markup can then be parsed to rebuild paragraphs, headings, and lists that were lost in the PDF's raw instruction stream.

A technique where optical character recognition is applied only to user-defined regions of a document, ignoring irrelevant areas like headers, footers, or marginalia. HOCR's coordinate system enables precise zonal targeting by allowing post-processing scripts to filter recognized text based on its bounding box position, effectively implementing zonal extraction after full-page recognition.

A token-level annotation standard where tokens are tagged as the Beginning, Inside, or Outside of a named entity. When building machine learning models for legal structure parsing, HOCR output is often tokenized and converted to BIO format to train sequence labeling models that identify clause boundaries, party names, and statutory references within the recognized text stream.