How to Translate Two-Column Academic PDFs While Keeping Equations and Layout

When translating dense two-column academic PDFs, text often scrambles across columns and math formulas break. Here is why the underlying PDF format fails, how layout-aware translation reconstructs reading order, and where university integrity rules draw the line.

HumanPen Team

· 9 min read

Why two-column academic PDFs break when translated and how to preserve layout

Two-column academic PDFs break during translation because the PDF standard stores visual glyph coordinates rather than semantic text flows. Naive extractors interleave columns, shatter hyphenated words, garble ligatures, and destroy mathematical formulas. Layout-preserving translation reconstructs geometric reading order and protects formula regions. Using translated papers for literature comprehension is widely accepted, but submitting machine-translated text as assessed coursework violates academic integrity rules across major universities.

Graduate students, researchers, and engineers routinely encounter this frustration when reading dense papers from IEEE, ACM, arXiv, or biomedical journals. Copying a paragraph into a web translation tool often merges sentences from the left and right columns horizontally, turns inline formulas into unreadable symbols, and splits compound terms across line breaks. Resolving these issues requires understanding why the underlying format behaves this way, how layout-aware engines parse geometric bounding boxes, and where university academic integrity policies draw the line between study assistance and unauthorized translation.

Inside a two-column PDF: glyph coordinates, TJ arrays, and untagged interleaving

To understand why translation tools fail on academic documents, one must examine how the Portable Document Format actually stores text. Governed by ISO 32000-1 (PDF 1.7), a PDF is essentially a visual printing instruction set, not a structured digital manuscript. As the technical documentation for open-source parsing libraries like `pdfminer.six` and `pypdf` explains, a PDF file does not contain native sentences, paragraphs, or word streams. Instead, each character is placed individually on an absolute two-dimensional canvas with specific bottom-left and upper-right coordinates.

"Most PDF files look like they contain well-structured text. But the reality is that a PDF file does not contain anything that resembles paragraphs, sentences or even words. When it comes to text, a PDF file is only aware of the characters and their placement."

This spatial representation creates immediate hurdles for text extraction. PDF specifications do not even define an explicit space character. In operators such as `TJ`, text is rendered as an array of character strings interleaved with numeric values that adjust horizontal kerning distances. Parsers must infer where words begin and end by calculating whether the horizontal gap between adjacent bounding boxes exceeds an arbitrary threshold, such as `word_margin`. If kerning values fluctuate or formulas introduce variable horizontal spacing, standard extractors either insert unintended spaces or fuse words together.

Column interleaving is the natural consequence of this design. As documented in W3C WCAG 2.2 Technique `PDF3`, reading order in a PDF depends on document structure tags. When an academic paper lacks complete tags, screen readers and naive extraction pipelines process text strictly from top to bottom across the entire page width. As `PyMuPDF` documentation highlights, applying a naive spatial sort (`sort=True`) simply orders content from top-left to bottom-right across the page canvas. Because line baselines in the left column align closely with those in the right column, alternating vertical coordinates cause text extractors to read line one of the left column, jump to line one of the right column, and scramble the narrative completely. Furthermore, elements inserted later in the document creation process, such as running headers or page numbers, are frequently dumped at the very end of extracted text streams.

The four technical failure modes: columns, hyphens, ligatures, and equations

When readers attempt to translate academic PDFs without layout-aware reconstruction, four distinct technical failures systematically corrupt the text. It is essential to distinguish native vector PDFs generated by LaTeX or Word from scanned image-only documents that require optical character recognition (OCR). Vector PDFs contain clean digital vector fonts and coordinate data, but their layout structures break down under naive extraction algorithms that treat 2D coordinate streams as linear text.

Failure ModeUnderlying PDF MechanismPlain Translation BehaviorLayout-Preserving Requirement
Column InterleavingLack of structural tags leads extractors to sort lines strictly by vertical coordinate across page margins.Sentences from left and right columns alternate line by line, producing incoherent paragraphs.Geometric layout analysis detects column gutters and isolates independent bounding box flows.
Broken End-of-Line HyphensWords split across lines with trailing hyphens (`-`) are treated as literal punctuation marks.Compound terms break into fragmented parts, generating mistranslated root words.Dehyphenation routines merge split spans before passing text to linguistic engines.
Garbled LigaturesTypographic combinations like `fi`, `fl`, and `ffi` are stored as single Unicode codepoints (such as `U+FB00`).Engines fail to decompose composite glyphs, leaving missing letters or odd symbols.Character mapping expands ligatures into constituent letters (`ffi` into `f`, `f`, `i`) prior to translation.
Corrupted Math EquationsTeX symbol fonts often lack complete `ToUnicode` CMaps; indices and fractions occupy distinct vertical coordinates.Math symbols turn into `U+FFFD` or raw IDs, and multi-line equations collapse into gibberish.Math block detection isolates equation regions as protected graphical or LaTeX objects.

Mathematical formulas represent the most acute technical vulnerability. Academic papers compiled in TeX frequently utilize subsetted math fonts that omit comprehensive `ToUnicode` CMap tables. As `pypdf` and `PyMuPDF` document, when an extractor encounters glyphs without Unicode mappings, it either falls back to replacement characters (`U+FFFD`) or exposes internal character IDs (`TEXT_USE_CID_FOR_UNKNOWN_UNICODE`). Furthermore, fractions, summations, superscripts, and subscripts are placed at varying vertical baselines rather than along a single linear baseline. When a naive translator receives these fragmented pieces, it attempts to translate mathematical variables as ordinary language, obliterating the scientific meaning of the paper.

Academic integrity boundaries: reading literature versus submitting coursework

While solving the engineering challenges of two-column PDF translation allows researchers to digest foreign-language papers efficiently, users must observe the critical boundary defined by university academic integrity regulations. Higher education institutions maintain a fundamental distinction between using machine translation as an assistive tool for literature comprehension and using translated text to produce assessed student coursework. For detailed discussions of how institutions evaluate AI tools across formats, see our analysis of translation tools under university rules and how Turnitin handles AI detection in PDF submissions.

"Writing your work in your native language and translating it into English before submission."

This boundary is articulated clearly across major academic frameworks. Newcastle University's official guidance on Using AI for Translation explicitly categorizes using translation tools to comprehend difficult assignments, define specialist technical terminology, and read foreign-language articles as acceptable practice, while categorizing drafting an essay in a native language and translating it into English for assessment submission as unacceptable misconduct. The Australian National University (ANU) Law School enforces a similar standard, explicitly prohibiting students from drafting assessments in another language and translating them via software, while mandating that any assistive translation used even for specific word clarifications be disclosed in the first footnote. Similarly, the University of Leeds (in its 2026/27 Academic Integrity and Assistance Policy v1.0) includes translation aids within digital technologies and holds students accountable for demonstrating genuine authorship, while the Beijing Technology and Business University Graduate School (in its March 2026 notice) permits translation assistance only with transparent disclosure and strictly forbids presenting machine output as original core analysis. Translating academic PDFs must remain an input-side learning aid, never an unacknowledged output-side drafting shortcut.

Translating academic PDFs while preserving columns, equations, and figures

Achieving an accurate translation without destroying the visual architecture of an academic paper requires a layout-aware processing pipeline. Instead of dumping raw characters into an unstructured translation buffer, a layout-preserving system first performs geometric document analysis to identify column boundaries, running headers, footnotes, vector figures, and mathematical equation blocks. Text blocks are extracted within their precise spatial boundaries, dehyphenated, and cleared of ligature anomalies before undergoing translation, while formula coordinates and figure containers remain isolated from linguistic replacement.

For researchers handling complex bilingual document workflows, `HumanPen` supports PDF, Word, PowerPoint, Excel, EPUB, HTML, and text between Chinese and English. Translate a document between Chinese and English, keeping the layout. The system is designed to preserve two-column layouts, figures, and math regions, though users should always verify output formatting and adhere to their institution's assessment policies before referencing translated materials.

  • Verify multi-column flow: Confirm that the translation engine followed the left column to its conclusion before reading into the right column, especially around floating figures or multi-column spanning tables.
  • Check math equation isolation: Inspect displayed equations and inline math symbols to verify that subscripts, Greek characters, and integration limits were preserved as intact mathematical entities rather than mistranslated as plain text.
  • Inspect container text expansion: Chinese and English feature different typographic densities, so review narrow tables and figure captions to ensure expanded translated sentences do not overflow their designated boxes.
  • Cross-reference critical citations: Compare numeric values, p-values, sample sizes, and bracketed citation numbers side-by-side against the original vector PDF to ensure no numerical figures were altered during translation.

Frequently asked questions

Why do copied formulas from an arXiv or IEEE PDF turn into random question marks or weird boxes? Mathematical expressions in papers compiled with TeX often use subsetted symbol fonts that lack comprehensive `ToUnicode` CMap tables. As documented by `pypdf` and `PyMuPDF`, when a PDF reader encounters a character code without a defined Unicode equivalent, it outputs a replacement glyph like `U+FFFD` or raw character identifiers. Furthermore, nested fractions, radicals, and super/subscripts are positioned at independent vertical coordinate baselines rather than as inline text, causing naive clipboards and translators to scatter the formula across multiple broken lines.

Can running OCR fix column interleaving in native vector PDFs? Optical character recognition is designed to extract text from scanned bitmap images, not native vector PDFs. Applying OCR to a digital vector PDF rasterizes crisp scalable vector fonts into low-resolution pixels, which frequently introduces spelling errors, degrades complex mathematical symbols, and still relies on external heuristics to distinguish two-column layouts. Native vector PDFs should instead be parsed using geometric layout analysis that clusters existing vector bounding boxes into distinct column containers without converting the document into an image.

Does translating an academic paper for reading violate university academic integrity policies? Using translation tools as an assistive study aid to understand foreign-language research articles, clarify unfamiliar technical terminology, or review literature generally complies with academic integrity guidelines at institutions like Newcastle University and the University of Leeds, provided students critically evaluate the translation and maintain independent academic judgment. Academic misconduct arises when a student drafts their own assessed coursework in their native language, uses automated translation to generate an English submission, and presents the resulting machine output as their own independent writing.

Why does translated Chinese text sometimes look unevenly spaced or clipped at column boundaries? Chinese and English typography obey fundamentally different spatial metrics. English text consists of variable-width proportional Latin glyphs that pack tightly in horizontal sentences, whereas Chinese characters are wider monospaced ideographs that require greater vertical line height and different word-wrapping boundaries. When English paragraphs are translated into Chinese within fixed-width PDF column containers, the resulting text can exceed container boundaries or display irregular character spacing unless the layout engine dynamically adjusts font sizes, line spacing, and box padding.

Sources

KEEP READING