Summary
On math-heavy single-column pages (gnuplot-rendered figures + equations) the extractor mis-orders inline math fragments and orphaned subscripts/superscripts. Numerals and superscripts emitted by \textsuperscript{} / _{...} ^{...} blocks land on their own short lines, separated from the base symbol.
Examples from https://arxiv.org/pdf/1601.03642.pdf:
expected: ... activation function ϕ is applied ...
actual: '... activation func- is blue.\n tion ϕ is applied. ...' (sub-caption interleaves)
expected: 1 / (1 + e^{-x})
actual: '1+1e−x' (figure-region equation collapsed)
expected: ... E with respect to some class of tasks T and ...
actual: 'wxxx... w0 \n x0 \n x1 \n w1 \n w2 \n Martin Thoma x2 Σ ϕ \n w \n ...'
The first page of 1601.03642 contains an artificial-neuron figure whose math glyphs (x_i, w_i, Σ, ϕ) are scattered single-per-line. Identical behaviour seen on 2201.00021 page 1 (NH₃ subscripts) and 2201.00200 page 1 (overshoot equations).
Reproduction
import pdf_oxide # 0.3.54
doc = pdf_oxide.PdfDocument("1601.03642.pdf")
print(doc.extract_text(0)[:600])Expected (ground truth)
Equations and figure-internal labels either grouped into a single coherent text block or omitted entirely.
Actual
Each subscript / superscript / single glyph emitted on its own line, interleaved with caption text of an adjacent figure region. Score on 1601.03642: 74.0% (markdown mode recovers to 93.0% by keeping the figure region contained).
Impact
- 1601.03642 (
creativity-in-ml.pdf by Thoma): -19 pt vs markdown mode
- 1602.06541 (
survey-of-semantic-segmentation.pdf): -16 pt vs markdown mode
- 2201.00200, 2201.00021 (equation-rich astro papers): -10 pt each
Hypothesis
The figure regions are detected as text regions by extract_text, and within them every glyph from the gnuplot/PSTricks figure becomes a separate "line" because the y-coordinate sort treats each glyph as its own row (figures often have lots of glyphs at slightly different y). The XY-cut pipeline (used in markdown) correctly clusters them into a single block.
Suggested fix
When extract_text runs, apply a coarse y-band clustering (tolerance ≥ 0.5 × line-height) before emitting glyphs in y-order. Alternatively, integrate the figure-detector that to_markdown_all evidently uses to skip image-region text entirely.
Related
Summary
On math-heavy single-column pages (gnuplot-rendered figures + equations) the extractor mis-orders inline math fragments and orphaned subscripts/superscripts. Numerals and superscripts emitted by
\textsuperscript{}/_{...}^{...} blocks land on their own short lines, separated from the base symbol.Examples from https://arxiv.org/pdf/1601.03642.pdf:
The first page of 1601.03642 contains an artificial-neuron figure whose math glyphs (
x_i,w_i,Σ,ϕ) are scattered single-per-line. Identical behaviour seen on 2201.00021 page 1 (NH₃ subscripts) and 2201.00200 page 1 (overshoot equations).Reproduction
Expected (ground truth)
Equations and figure-internal labels either grouped into a single coherent text block or omitted entirely.
Actual
Each subscript / superscript / single glyph emitted on its own line, interleaved with caption text of an adjacent figure region. Score on 1601.03642: 74.0% (markdown mode recovers to 93.0% by keeping the figure region contained).
Impact
creativity-in-ml.pdfby Thoma): -19 pt vs markdown modesurvey-of-semantic-segmentation.pdf): -16 pt vs markdown modeHypothesis
The figure regions are detected as text regions by
extract_text, and within them every glyph from the gnuplot/PSTricks figure becomes a separate "line" because the y-coordinate sort treats each glyph as its own row (figures often have lots of glyphs at slightly different y). The XY-cut pipeline (used in markdown) correctly clusters them into a single block.Suggested fix
When
extract_textruns, apply a coarse y-band clustering (tolerance ≥ 0.5 × line-height) before emitting glyphs in y-order. Alternatively, integrate the figure-detector thatto_markdown_allevidently uses to skip image-region text entirely.Related