Font-Based Heuristic Parsing

The primary mechanism for detecting heading levels by establishing numeric thresholds for font point sizes. A parser scans the document's text runs and compares each run's size against predefined or dynamically calculated breakpoints.

• Absolute Thresholding: Classifies text above 14pt as a heading, below as body text.
• Relative Thresholding: Identifies the largest font on a page as the primary heading and scales subsequent levels proportionally.
• Modal Analysis: Computes the most frequent font size to establish the body text baseline, then flags any text exceeding this baseline by a set delta as a structural element.

Parsers analyze the bold and italic properties of text runs to disambiguate structural roles when font size alone is ambiguous. A bold 12pt run might be a subheading, while a regular 12pt run is body text.

• Bold Detection: Checks the font-weight property (values ≥700 in CSS or a dedicated bold font file reference in PDFs).
• Italic Analysis: Often signals a blockquote, caption, or procedural history section in legal documents.
• Combined Heuristics: A rule might state: 'If font size > body baseline AND weight = bold, classify as Level 2 Heading.'

Changes in typeface family (e.g., switching from Times New Roman to Arial) often signal a semantic shift in legal documents, such as a move from the main body to a block quotation or an appendix.

• Mono-Signal Detection: A sudden switch to a monospaced font like Courier New strongly indicates a verbatim clause or a statutory excerpt.
• Sans-Serif Shifts: In contracts, a shift to a sans-serif font often marks a defined term or a header.
• Fallback Logic: If the font family is not extractable, the parser falls back to glyph width analysis to infer the typeface category.

Font heuristics are combined with layout geometry to validate structural inferences. A large, bold text block is only a heading if it precedes a body text block with standard indentation.

• Left-Alignment Check: Headings often have a distinct left indent compared to body paragraphs.
• Vertical Gap Analysis: Measures the white space above and below a text run. A larger gap above than below signals a section start.
• Orphan Prevention: Ensures a single bold line at the bottom of a page is not misclassified as a heading if the following text is on the next page.

Font heuristics are cross-referenced with outline numbering patterns (e.g., 'I.', 'A.', '1.', 'a.') to confirm the hierarchy. A bold 12pt line starting with '(a)' is a sub-subsection, not a main heading.

• Regex Matching: Scans for patterns like ^[IVX]+. or ^\d+\. at the start of a text run.
• Romanet Detection: Specifically identifies lowercase Roman numerals (i, ii, iii) used in deep nesting.
• Hierarchy Validation: If a parser detects a 'Section 1.2' heading, it verifies that a 'Section 1.1' was previously identified to maintain logical sequence integrity.

When font metadata is stripped or uniform (e.g., in a plain-text brief), the parser falls back to linguistic patterns to infer structure. This hybrid approach bridges the gap between visual and semantic parsing.

• Keyword Triggers: Lines starting with 'WHEREAS', 'NOW THEREFORE', or 'ARTICLE' are classified as structural boundaries regardless of font.
• Sentence Length Analysis: Headings are typically shorter than 15 words; body text is longer.
• Punctuation Cues: Headings rarely end with a period, while body text sentences do. This simple rule resolves ambiguity in poorly formatted documents.