Academic papers combine nearly every difficult PDF feature: two-column layouts, formulas, tables, footnotes, figure captions, references, and publisher-specific fonts. A clean title and abstract can create false confidence while the methods section is interleaved and equations are silently damaged.
The right workflow depends on the purpose. Reading notes need understandable prose. A citation database needs accurate metadata. A RAG corpus needs section structure and provenance. Republishing requires attention to copyright and accessibility. Define the goal before choosing a parser.
Begin with document rights and provenance
Possessing a PDF does not automatically grant permission to republish its full text. Personal conversion for reading or analysis differs from placing an article online. Retain the citation, DOI, authors, venue, publication date, and source URL. Do not remove license notices.
Record the exact source version. Preprints, accepted manuscripts, and versions of record can differ. If the Markdown enters a dataset, include a checksum and conversion date so later users can identify the underlying file.
The Markdown is a derivative. Keep the source available under appropriate access controls until review is complete.
Test reading order early
Most papers use two columns after a full-width title and abstract. Basic coordinate sorting may alternate lines between columns. Convert a body page and confirm that the left column finishes before the right begins. Check pages containing figures because captions and floating boxes disrupt simple column boundaries.
Section headings may span one column or the full page. Verify that the introduction, methods, results, and discussion remain in sequence. Running headers, page numbers, and journal branding should usually be removed from the prose.
If several test pages fail, use a layout-aware parser rather than manually repairing an entire corpus. The multi-column guide explains the decision in more detail.
Treat equations as a separate requirement
Equations may be real text, embedded fonts, vector paths, or images. A simple extractor can return recognizable variables while losing fractions, superscripts, subscripts, matrices, and alignment. The result may look plausible but be mathematically wrong.
Decide whether equations need to remain executable LaTeX, readable images, or omission markers. If mathematical fidelity matters, use a specialized model and compare every important formula with the source. Do not convert x² to x2 or drop a negative sign without flagging the defect.
Inline symbols also deserve review. Greek letters, arrows, multiplication signs, and Unicode normalization can change meaning. Preserve equation numbers when the prose references them.
Reconstruct tables and figures carefully
Scientific tables often have grouped headers, significance markers, units, and footnotes. Markdown tables cannot represent every relationship. Validate row and column labels, decimals, minus signs, confidence intervals, and units. Use CSV or HTML when that preserves structure better.
Text extraction generally does not include figure images. Retain captions with a clear placeholder, and export figures separately only when rights and purpose permit. Alt text should describe the information, not merely say “Figure 2.”
Ensure captions remain next to the correct figure marker. Floating placement can cause a caption to appear inside unrelated prose.
Preserve citations and references
In-text citations may be numeric, author-year, superscript, or linked annotations. Check bracket pairs and ranges such as [4–7]. A line break can split a citation or detach it from the sentence.
Reference lists have hanging indents and many punctuation patterns. A layout parser may mix entries when a reference wraps across lines. Use DOI and bibliographic metadata services to validate identifiers, but do not replace the source citation silently. Record corrections.
For knowledge systems, store references as structured metadata when possible. A Markdown bibliography is readable but difficult to deduplicate reliably.
Handle footnotes and supplementary material
Footnotes can appear at the bottom of a column, page, or article. Keep markers connected to notes and distinguish author affiliations from explanatory footnotes. Markdown footnote syntax is useful if the target renderer supports it; otherwise use a consistent labeled section.
Supplementary appendices may have a different layout or page numbering. Confirm whether they are inside the same PDF and whether they are required for the intended analysis. A successful main article does not mean supplementary tables were converted.
Choose tools by document difficulty
A browser text-layer converter is appropriate for a clean, mostly linear paper when privacy and zero setup matter. It is useful for quickly extracting prose from a single-column preprint or reading the abstract and discussion.
Marker and MinerU are stronger candidates for formulas, scans, dense tables, and complex layouts. They require local installation, model downloads, and more compute. Cloud scholarly parsers may add metadata and references, but document handling must fit your privacy and licensing constraints.
Run the same representative pages through candidate tools. Compare the structures that matter to your task instead of relying on a global accuracy claim.
Review according to downstream use
For personal notes, verify section order and quoted passages. For systematic review, validate metadata, inclusion criteria, numbers, and citations. For RAG, clean headers, preserve heading paths, and attach page provenance. For publication, address rights, figures, accessibility, and full editorial review.
Search the Markdown for replacement characters, isolated page numbers, suspicious equation fragments, and broken references. Render it and inspect heading hierarchy. Sample beginning, middle, complex, and final pages.
The RAG preparation guide covers ingestion design, while the cleanup checklist provides a general production pass. Academic conversion succeeds when limitations are visible and evidence remains traceable—not when the output merely contains many words.
