# Dataset & Benchmark Results

---

# Experimental Benchmark & Evaluation

## 1. Archival Corpus

Evaluated across **609 text segments** spanning **45 digitized pages** from 9 Swiss French-language historical periodicals (Canton of Vaud, 1733–1945):

<table id="bkmrk-periodical-era-issue"> <thead> <tr> <th style="text-align:left;">Periodical</th> <th style="text-align:left;">Era</th> <th style="text-align:left;">Issues</th> <th style="text-align:left;">Pages</th> <th style="text-align:left;">Segments</th> <th style="text-align:left;">Characteristics</th> </tr> </thead> <tbody> <tr> <td>**La Revue**</td> <td>1875–1945</td> <td>4</td> <td>5</td> <td>139</td> <td>19th/20th century serif, multi-column</td> </tr> <tr> <td>**Feuille d'Avis**</td> <td>1762–1841</td> <td>4</td> <td>6</td> <td>118</td> <td>Antique typography, long 's' ligatures</td> </tr> <tr> <td>**Mercure Suisse**</td> <td>1733–1748</td> <td>3</td> <td>5</td> <td>92</td> <td>18th-century early modern French</td> </tr> <tr> <td>**Gazette de Lausanne**</td> <td>1804–1920</td> <td>5</td> <td>8</td> <td>145</td> <td>High density, complex layout</td> </tr> <tr> <td>**Others (Tribune, etc.)**</td> <td>1880–1930</td> <td>5</td> <td>21</td> <td>115</td> <td>Mixed commercial &amp; political press</td> </tr> </tbody></table>

---

## 2. Key Benchmark Results

### Error Rate Comparison (Tesseract OCR Base):

- **Raw OCR Baseline:** *CER = 6.30\\%*, *WER = 14.20\\%*
- **All-LLM Baseline (GPT-4o / Gemini 1.5 Flash):** *CER = 2.92\\%* (Cost: 100% token budget)
- **Selective Pre-Routing (Ours):** *CER = 3.12\\%* (Bypasses **40%** of all segments!)
- **Collaborative Safeguard (&lt;5% Human Budget):** ***CER = 2.49\\%*** (**14% relative reduction** over All-LLM)

### Comparison with Routing Baselines:

1. **Random Routing:** Weakest; frequently sends clean text and degrades it.
2. **OCR Confidence (Thresholding):** Underperforms because Tesseract confidence scores poorly correlate with semantic error in archaic fonts.
3. **ConfBERT:** Good detection but computationally heavy.
4. **Our Regression Router + Safeguard:** Consistently establishes the Pareto frontier across all budget limits.

---

## Experimental Evaluation &amp; Visual Benchmark Plots

### Historical Newspaper Segmentation Sample

<div id="bkmrk-figure%3A-doceng_segme" style="text-align:center;"><div style="text-align:center;">[![](https://wiki.stergios.ch/uploads/images/gallery/2026-09/scaled-1680-/doceng-segment-example.png)](https://wiki.stergios.ch/uploads/images/gallery/2026-09/doceng-segment-example.png "Click to view full-resolution image")</div>*Figure 1: Representative historical archival newspaper segment illustrating layout artifacts and OCR degradation.*

</div>### CER Reduction Across Correction Strategies

<div id="bkmrk-figure%3A-correction_s" style="text-align:center;"><div style="text-align:center;">[![](https://wiki.stergios.ch/uploads/images/gallery/2026-09/scaled-1680-/correction-strategies-comparison.png)](https://wiki.stergios.ch/uploads/images/gallery/2026-09/correction-strategies-comparison.png "Click to view full-resolution image")</div>*Figure 2: Character Error Rate (CER) reduction across baseline OCR, All-LLM, ConfBERT, and Collaborative Routing.*

</div>### OCR Engine Improvement Comparison

<div id="bkmrk-figure%3A-engine_impro" style="text-align:center;"><div style="text-align:center;">[![](https://wiki.stergios.ch/uploads/images/gallery/2026-09/scaled-1680-/engine-improvement-chart.png)](https://wiki.stergios.ch/uploads/images/gallery/2026-09/engine-improvement-chart.png "Click to view full-resolution image")</div>*Figure 3: Relative error reduction across historical OCR engines (Tesseract, Abbyy, and modern HTR).*

</div>### Strategy Performance Heatmap

<div id="bkmrk-figure%3A-engine_strat" style="text-align:center;"><div style="text-align:center;">[![](https://wiki.stergios.ch/uploads/images/gallery/2026-09/scaled-1680-/engine-strategy-heatmap.png)](https://wiki.stergios.ch/uploads/images/gallery/2026-09/engine-strategy-heatmap.png "Click to view full-resolution image")</div>*Figure 4: Comprehensive performance heatmap comparing LLM prompt strategies across historical periodicals.*

</div>