Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CNN-based Image Models Verify a Hypothesis that The Writers of Cuneiform Texts Improved Their Writing Skills When Studying at the Age of Hittite Empire

Daichi Kohmoto, Katsutoshi Fukuda, Daisuke Yoshida, Takafumi Matsui, Sachihiro Omura

TL;DR

This paper investigates whether the two authors of the cuneiform tablet KBo 23.1 ++/KUB 30.38 show a teacher–student dynamic and skill improvement, using a data-driven CNN-based image analysis that avoids character-by-character segmentation. The authors build 40 main datasets from three tablet images, fine-tune three CNNs, define a class-similarity metric from confusion matrices, and employ a 2-step majority vote along with CAM-based explainability to infer authorship and progression. They find evidence supporting the two-author hypothesis, with the latter author appearing to improve writing through practice, and demonstrate the method’s applicability to additional tablet fragments via external validation. The work offers a simple, extensible approach for handwriting analysis and writer identification in archaeology and historical linguistics, highlighting how image-based methods can reveal pedagogical practices in ancient writing traditions.

Abstract

A cuneiform tablet KBo 23.1 ++/KUB 30.38, which is known to represent a text of Kizzuwatna rituals, was written by two writers with almost identical content in two iterations. Unlike other cuneiform tablets that contained information such as myths, essays, or business records, the reason why ancient people left such tablets for posterity remains unclear. To study this problem, we develop a new methodology by analyzing images of a tablet quantitatively using CNN (Convolutional Neural Network)-based image models, without segmenting cuneiforms one-by-one. Our data-driven methodology implies that the writer writing the first half was a `teacher' and the other writer was a `student' who was training his skills of writing cuneiforms. This result has not been reached by classical linguistics. We also discuss related conclusions and possible further directions for applying our method and its generalizations.

CNN-based Image Models Verify a Hypothesis that The Writers of Cuneiform Texts Improved Their Writing Skills When Studying at the Age of Hittite Empire

TL;DR

This paper investigates whether the two authors of the cuneiform tablet KBo 23.1 ++/KUB 30.38 show a teacher–student dynamic and skill improvement, using a data-driven CNN-based image analysis that avoids character-by-character segmentation. The authors build 40 main datasets from three tablet images, fine-tune three CNNs, define a class-similarity metric from confusion matrices, and employ a 2-step majority vote along with CAM-based explainability to infer authorship and progression. They find evidence supporting the two-author hypothesis, with the latter author appearing to improve writing through practice, and demonstrate the method’s applicability to additional tablet fragments via external validation. The work offers a simple, extensible approach for handwriting analysis and writer identification in archaeology and historical linguistics, highlighting how image-based methods can reveal pedagogical practices in ancient writing traditions.

Abstract

A cuneiform tablet KBo 23.1 ++/KUB 30.38, which is known to represent a text of Kizzuwatna rituals, was written by two writers with almost identical content in two iterations. Unlike other cuneiform tablets that contained information such as myths, essays, or business records, the reason why ancient people left such tablets for posterity remains unclear. To study this problem, we develop a new methodology by analyzing images of a tablet quantitatively using CNN (Convolutional Neural Network)-based image models, without segmenting cuneiforms one-by-one. Our data-driven methodology implies that the writer writing the first half was a `teacher' and the other writer was a `student' who was training his skills of writing cuneiforms. This result has not been reached by classical linguistics. We also discuss related conclusions and possible further directions for applying our method and its generalizations.
Paper Structure (17 sections, 8 equations, 9 figures, 5 tables)

This paper contains 17 sections, 8 equations, 9 figures, 5 tables.

Table of Contents

  1. Preface
  2. Methodology and Experiments
  3. Building Main Datasets
  4. Study of Similarity between Classes
  5. Study With a Majority Vote
  6. Fine-Tuning CNN-based Image Models
  7. Class Activation Mapping (CAM): Explainability of Models
  8. Demonstrating Fine-Tuned Models for Other Cuneiform Sentences
  9. Results and Discussion
  10. Learning Curves
  11. Confusion Matrices
  12. Class Activation Mapping
  13. Demonstrating Fine-Tuned Models to Other Cuneiform Sentences
  14. Conclusion
  15. Supplementary information
  16. ...and 2 more sections

Figures (9)

  • Figure 1: Raw Images of Target Tablet KBo 23.1 ++/KUB 30.38: (a) Left-Bottom Part of Front Side 1, (b) Right-Top Part in Front Side 2, and (c) Back Side
  • Figure 2: Definition of Classes: (a) Class $1$ in $4$ classes, (b) Class $2$ in $4$ classes, (c) Class $3$ & $4$ in $4$ classes, (d) Class $1,2$ in $8$ classes, (e) Class $3,4$ in $8$ classes, and (f) Class $5$-$8$ in $8$ classes
  • Figure 3: Cuneiform Sentences for Studying Their Authors: (a) Right-Bottom Part of Front Side and (b) The Sides of Tablet
  • Figure 4: Learning Curves for Fine-Tuned Image Models on v001 Dataset, the Case of Seed $1033$: (a) VGG19, (b) ResNet50, and (c) InceptionV3
  • Figure 5: Overall Confusion Matrices via Fine-Tuned Models, for $4$ Classes: (a) VGG19, (b) ResNet50, and (c) InceptionV3. C$1$-C$4$ stand for Class $1$-$4$, respectively
  • ...and 4 more figures