Quantifying Page Segmentation Quality in Historical Job Advertisements Retrieval - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2024

Quantifying Page Segmentation Quality in Historical Job Advertisements Retrieval

Résumé

This paper addresses the question of which metric is the most suitable for evaluating page segmentation in the context of extracting historical job advertisements in digitized newspapers. Accurate page segmentation is essential for high-quality Optical Character Recognition (OCR) results, yet the methodology for comparing and evaluating segmentation algorithms has received limited attention in Digital Humanities. The paper presents an evaluation framework developed within the JobAds Project, focusing on textual congruence between predicted and ground-truth regions. This is important for an evidence-based selection of the best-performing segmentation algorithm, and offers insights into the resulting segmented data quality, which in turn impacts research outcomes. The paper examines three evaluation features: intersection area, text similarity based on Levenshtein distance, and text presence/absence in non-intersecting parts of the predicted region and its ground truth, revealing their effectiveness through logistic regression models. The method involves manual ground-truth creation, aiming for an automatic metric to quantify textual congruence. Results show that combining the text presence/absence feature with Hausdorff distance achieves the highest performance, reaching an F1 score of 0.957 on the testing subset. The study emphasizes the need for tailored evaluation metrics in Digital Humanities according to the specific needs and goals. The proposed evaluation framework offers insights for segmentation assessment in historical newspapers, with further application beyond the specific dataset and use case.
Fichier principal
Vignette du fichier
Segmentation Evaluation - final.pdf (1.87 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04560463 , version 1 (26-04-2024)

Licence

Identifiants

  • HAL Id : hal-04560463 , version 1

Citer

Klara Venglarova, Raven Adam, Saranya Balasubramanian, Georg Vogeler. Quantifying Page Segmentation Quality in Historical Job Advertisements Retrieval. 2024. ⟨hal-04560463⟩
100 Consultations
13 Téléchargements

Partager

Gmail Mastodon Facebook X LinkedIn More