Source-preserving alignment for robust evidence localization in scientific PDFS

1. INTRODUCTION

• Evidence strings must be matched to PDF text despite line breaks, Unicode variants, superscripts, citation markers, and fragmented items.These differences arise from the PDF text model and character encoding.Source: §1 · p. 1
• Text search is sensitive to normalization differences, while precomputed bounding boxes cannot handle unexpected text and are sensitive to layout or encoding changes.Neither baseline preserves the character-level link to source geometry needed for accurate highlighting.Source: §1 · p. 1
• The framework aligns normalized text, source characters, and page geometry, mapping exact or approximate matches to source spans for rendering.Its central principle is to normalize for matching while preserving provenance for localization.Source: §1 · p. 1

2. RELATED WORK

• PDF libraries expose page strings, text items, and geometry, while neural parsers improve structured extraction without guaranteeing traceable visual source spans.Standard search is fragile to extraction-order, encoding, and layout differences; bounding-box indexes are limited by text-to-box correspondence.Source: §2 · p. 1
• Scientific PDFs fragment or encode equations, formulas, Greek letters, superscripts, and reference markers nonstandardly, amid substantial layout variation.A robust locator therefore needs tolerant comparison together with faithful recovery of source characters and geometry.Source: §2 · pp. 1–2
• Approximate matching, sequence alignment, and LCS identify correspondences despite textual errors or gaps [16–18].This paper's token-level LCS fallback projects supported tokens through a provenance map to source characters and page geometry.Source: §2 · p. 2

3.1. Deterministic Matching Procedure

• Fixed rules search for an exact occurrence first, then enumerate candidate windows for fallback matching when exact matching fails.The fallback is accepted only at score ≥0.8; otherwise, no fallback match is returned.Source: §3.1 · p. 2

3.2. Normalization and Provenance Mapping

• Normalization unifies Unicode and dash variants, repairs malformed encoding, maps scientific symbols, and removes layout-only separators.Each normalized character retains its source item and character offset, allowing matches to project back to source intervals; original page text is not rewritten.Source: §3.2 · p. 2

3.3. Exact Matching and Character-Range Clipping

• For exact matches, the method renders only matched character intervals and clips the transformed rectangle to the page viewport.Horizontal bounds are estimated proportionally from item width, preventing a quote ending mid-item from highlighting the following sentence.Source: §3.3 · p. 2

3.4. Line-Break and Noisy-Text Alignment

• When exact matching fails, token windows at several lengths are compared using LCS; only aligned tokens are highlighted, excluding unaligned markers.Alphabetic line-end hyphenation is merged when fragments form a word, while formula-like strings such as “ZSM-5” are preserved.Source: §3.4 · p. 2

3.5. Geometric Rendering

• Rendering transforms text-rectangle corners and merges nearby fragments only on the same rendered line.This avoids merging across columns, headers, or unrelated text regions.Source: §3.5 · p. 2

4.1. Dataset

• The benchmark contains 1,020 chemistry papers and 18,580 quote instances, averaging 18.2 quotes per paper.Only papers with both PDF and Markdown files were retained, so the benchmark evaluates localization rather than upstream claim extraction.Source: §4.1 · p. 2
• The corpus includes varied quote lengths, multi-column pages, chemical formulas, figure references, and noisy text layers.These features expose ordinary search failures and boundary errors after apparently successful matches.Source: §4.1 · p. 2

4.2. Comparison Methods

• The comparison evaluates the proposed source-preserving method against PyMuPDF direct lookup and an offline page-text and bounding-box index.The proposed method combines source-preserving normalization, alignment, source-span recovery, and geometry-aware rendering.Source: §4.2 · p. 3

4.3. Evaluation Metrics and Human Verification Protocol

• Manual verification covers 18,572 quotes per method, with Correct requiring fully accurate highlighting and Incorrect including boundary excess or omission.The remaining eight of 18,580 quotes have no manual labels.Source: §4.3 · p. 3

4.4. Overall Performance

• 92.6% automatic localization rate for the proposed method exceeds 43.6% for text search and 19.1% for precomputed BBOX.Automatic localization measures whether a method returns at least one candidate region for a quote.Source: §4.4 · p. 3, §4.3 · p. 3

4.5. Component Ablation

• Normalization provides a larger localization-rate gain than approximate alignment, which further recovers disrupted token sequences.The ablation removes each component separately on the same 1,020 papers and 18,580 quotes.Source: §4.5 · p. 3, Figure 2 · p. 3
• Whole-item rendering adds 62.8 source characters per quote on average, while character-range clipping introduces no whole-item excess.This rendering-only comparison holds matched source spans fixed across 17,213 successfully localized quotes and does not change localization decisions.Source: §4.5 · p. 3

4.6. Robustness to Evidence Length

• The proposed method peaks at 81–160 characters, then declines but remains substantially ahead of both baselines for longer quotes.BBOX slightly exceeds the proposed method in the shortest group.Source: §4.6 · p. 3, Figure 3 · p. 4
• Longer evidence can cross text items and line breaks or contain marker insertions and encoding variation; provenance mapping excludes unmatched content from highlights.PyMuPDF peaks at 41–80 characters before declining, while BBOX declines across all groups.Source: §4.6 · pp. 3–4

4.7. Human Evaluation

• Six evaluators randomly assessed quote–method outputs by comparing each highlight with its evidence in the original PDF.Correct required exact location and boundaries; any positional error, excess, or omission was Incorrect, and no usable localization was Not found.Source: §4.7 · p. 4
• The proposed method achieves the highest Correct rate, while both baselines have larger Not found shares.Figure 4 summarizes the three-category results.Source: §4.7 · p. 4

• Residual errors mainly reflect disrupted extraction order in complex layouts, approximate character widths, or extraction errors beyond the normalization rules and LCS threshold.Source: §4.8 · p. 4
• The framework assumes that the PDF has a usable text layer.Source: §4.8 · p. 4
• Character-level provenance crops highlights at the final evidence character, excluding unsupported text that whole-item highlighting would include.Source: §4.8 · p. 4

5. LIMITATIONS AND FUTURE WORK

• Variable-width fonts can produce loose highlight boundaries because character geometry is estimated from item width.Source: §5 · p. 4
• Glyph-level metrics or native text-layer positioning could address the character-geometry limitation.Source: §5 · p. 4
• Layout-aware analysis, confidence calibration, and lightweight correction are proposed to improve reading order and support large-scale verification.Source: §5 · p. 4

6. CONCLUSION

• The framework maps normalized evidence through source characters to page geometry for robust matching and precise highlights.Source: §6 · p. 4
• The method substantially improves over text-search and precomputed-BBOX baselines, especially for long evidence strings.Source: §6 · p. 4
• Ablation results show that both normalization and approximate token alignment are necessary for the reported gain.Source: §6 · p. 4