Hybrid Multi-Phase Page Matching and Multi-Layer Diff Detection for Japanese Building Permit Document Review
A new hybrid algorithm achieves perfect precision and 80% recall in matching complex document revisions.
Researcher Mitsumasa Wada has published a paper detailing a novel AI system designed to automate the complex review of Japanese building permit documents. The system tackles a major pain point in construction regulation: manually cross-referencing large PDF document sets across multiple revision cycles is notoriously labor-intensive and prone to human error. Wada's solution is a hybrid, multi-phase algorithm that first intelligently matches pages between document versions, even when page order, numbering, or content changes substantially. It does this by combining a longest common subsequence (LCS) method for structural alignment, a sophisticated seven-phase consensus matching pipeline, and a final dynamic programming stage for optimal alignment.
Once pages are accurately paired, a second, multi-layer 'diff' engine analyzes the matched content to identify changes. This engine operates on three distinct levels: text-level for wording alterations, table-level for structural data changes, and pixel-level visual differencing for graphical or layout modifications. The combined output is a comprehensive, highlighted difference report. In evaluation on real-world permit document sets, the system demonstrated robust performance, achieving an F1 score of 0.80 and a perfect precision score of 1.00 on a manually annotated benchmark. Critically, it produced zero false-positive matched page pairs, indicating high reliability for practical deployment. This research, categorized under both Computation and Language (cs.CL) and Computer Vision (cs.CV), represents a significant step in applying multi-modal AI to solve specific, high-stakes bureaucratic and regulatory challenges.
- The hybrid algorithm uses a seven-phase consensus pipeline and dynamic programming to match pages across document revisions with 100% precision and zero false positives.
- A multi-layer diff engine performs text, table, and pixel-level visual analysis to produce comprehensive change reports.
- Evaluation on real-world Japanese building permit documents achieved an F1 score of 0.80, automating a manual, error-prone regulatory process.
Why It Matters
This automates a critical but tedious regulatory task, reducing human error and accelerating construction permit approvals in a major economy.