--- title: "Advanced Deep Learning Approaches for Automated Recognition of Cuneiform Symbols — Elshehaby 2025" description: "The authors present the work as a step toward less manual decipherment. 18 references, cited by 1 paper here." authors: ["Shahad Elshehaby", "Alavikunhu Panthakkan", "Hussain Al-Ahmad", "Mina Al-Saad"] year: 2025 venue: "arXiv:2505.04678" source: "https://arxiv.org/abs/2505.04678" doi: "10.48550/arxiv.2505.04678" arxiv: "2505.04678" licence: "CC BY-NC-ND 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by-nc-nd/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/elshehaby-2025-advanced-deep-learning-approaches-automated/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/elshehaby-2025-advanced-deep-learning-approaches-automated/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/2/) / Advanced Deep Learning Approaches for Automated Recognition of Cuneiform Symbols # Advanced Deep Learning Approaches for Automated Recognition of Cuneiform Symbols **Authors:** Shahad Elshehaby, Alavikunhu Panthakkan, Hussain Al-Ahmad and Mina Al-Saad · **Year:** 2025 · **Venue:** arXiv:2505.04678 · **Licence:** [CC BY-NC-ND 4.0 (arXiv)](http://creativecommons.org/licenses/by-nc-nd/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2505.04678) · [PDF](https://arxiv.org/pdf/2505.04678) · [DOI 10.48550/arxiv.2505.04678](https://doi.org/10.48550/arxiv.2505.04678) · [arXiv 2505.04678](https://arxiv.org/abs/2505.04678) · [OpenAlex](https://openalex.org/W4408727648) **Citations:** 3 citations (OpenAlex, 2026-10-14) · 1 in our library, of the 140,884 papers whose bibliographies we read ## Summary The paper tackles automatic reading of cuneiform signs. Interpreting them by hand is slow, tablets are often damaged, and a single sign can carry several meanings. The authors built an image dataset of 14,100 samples starting from 235 distinct signs taken from *A Manual of Akkadian* (Marcus). Each sign was cleaned up (standardized, resized, denoised) into 10 variants, and each variant then received five augmented copies. They trained five classifiers on this data: VGG16, EfficientNet(V2M), MobileNet, InceptionResNetV2 and a custom 2D CNN. Training used Adam, up to 50 epochs, and early stopping with a patience of five. EfficientNet and VGG16 had the lowest loss and the best test scores, both close to perfect. These two were then applied to the signs of Hammurabi Law 1, where they recovered the Akkadian meaning of each sign and an English rendering. A separate qualitative section relates Akkadian to Arabic as Semitic/Afro-Asiatic relatives. The authors present the work as a step toward less manual decipherment. They plan ensembles and stacking next, and possibly an extension to Egyptian hieroglyphs. ## Key points - Dataset: 235 sign classes × 10 preprocessed variants × (1 original + 5 augmentations) = 14,100 images. - The split is unusual: 40% test (5,640), 24% validation (3,384) and 36% train (5,076). - Models compared: VGG16, EfficientNetV2M, MobileNet, InceptionResNetV2 and a custom 2D CNN. - Training used Adam for 50 epochs, with early stopping (patience = 5) on the validation set. - Reported accuracy and F1 by model: | Model | Accuracy | F1 | |---|---|---| | EfficientNet | 0.9999 | 0.9999 | | VGG16 | 0.9994 | 0.9996 | | MobileNet | 0.9991 | — | | 2D CNN | 0.9875 | — | | InceptionResNetV2 | 0.9857 | — | - Caveats on the numbers: - VGG16's reported F1 is higher than both its precision and its recall. That cannot happen for a harmonic mean, so the table needs checking. - The paper does not say whether augmented copies of one sign are kept within a single split. Scores this close to perfect therefore deserve a leakage check. - The real-world test is a single text, Hammurabi Law 1. It is used for sign → Akkadian meaning → English output, and no quantitative score is given for it. ## Interesting ideas - Shortlist candidate models by how they behave in training and on validation, then send only the top two to an authentic, out-of-distribution document instead of trusting the synthetic test split alone. - The pipeline does not stop at a class label. Each recognized sign is mapped through a lexicon to a reading and then a gloss, so every output step can be inspected. - A known related language (Arabic) is used as a qualitative cross-check on whether the recovered meanings of the ancient language are plausible. - A handful of canonical samples per class are expanded into many variants through normalization plus augmentation, giving the model several views of each item. - The split reserves a larger share for testing than for training, putting reliable evaluation ahead of training volume. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (learn): after benchmark evaluation, run the top parser or teacher candidates end to end on a few genuine fables or school texts as a final gate before promotion. - **I2** (world): keep the chain from token to lexical entry to ontology concept explicit and stored, so any answer can be traced back to the sign or word that produced it. - **I3** (back): when translating through the meta-world, check outputs against cognates or close-language equivalents to catch implausible renderings. - **I4** (tokens): generate controlled spelling, casing and hyphenation variants of rare words to make the tokenizer and tagger robust without new annotation. - **I5** (learn): hold out large evaluation sets, including negative controls, when silver data is cheap, so improvements in repair rules are measured with tight confidence. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) Topics: [𒀭 Akkadian and cuneiform NLP](https://md.research.elamur.ai/free/en/topics/akkadian-cuneiform-nlp/) · [🐪 Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) ## Institutions Also on the byline: [University of Dubai](https://ror.org/05h0z7c09). ## Cite Cite as: Shahad Elshehaby, Alavikunhu Panthakkan, Hussain Al-Ahmad and Mina Al-Saad. 2025. Advanced Deep Learning Approaches for Automated Recognition of Cuneiform Symbols. arXiv:2505.04678. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{elshehaby2025advanced, title = {{Advanced Deep Learning Approaches for Automated Recognition of Cuneiform Symbols}}, author = {Shahad Elshehaby, Alavikunhu Panthakkan, Hussain Al-Ahmad and Mina Al-Saad}, year = {2025}, doi = {10.48550/arxiv.2505.04678}, eprint = {2505.04678}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2505.04678} } ``` - [← previous Parsing the Switch: LLM-Based UD Annotation for Complex Code-Switched and Low-Resource Languages](https://md.research.elamur.ai/free/en/papers/kellert-2025-parsing-switch-llm-based-ud/) - [next → An Ensemble Classification Approach in A Multi-Layered Large Language Model Framework for Disease Prediction](https://md.research.elamur.ai/free/en/papers/hamdi-2025-ensemble-classification-approach-multi-layered/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/2/) ## References 18 references, 0 of them with a card of ours. ↗ leads to the work. 1. J.-J. Glassner, Z. Bahrani, and M. Van De Mieroop, “The invention of cuneiform: writing in Sumer,” Choice Reviews Online, vol. 41, no. 11, pp. 41–6691, Jul. 2004, doi: 10.5860/choice.41-6691. [↗](https://doi.org/10.5860/choice.41-6691) 2. N. Mahadevaswamy, “The Code of Hammurabi - perhaps the first law code,” Journal of Emerging Technologies and Innovative Research, Nov. 2018, [Online]. Available: [https://www.jetir.org/view?paper=JETIR1811B49](https://www.jetir.org/view?paper=JETIR1811B49) [↗](https://www.jetir.org/view?paper=JETIR1811B49) 3. E. E. Knudsen and E. Reiner, “A Linguistic analysis of Akkadian,” Journal of the American Oriental Society, vol. 90, no. 2, p. 334, Apr. 1970, doi: 10.2307/598174. [↗](https://doi.org/10.2307/598174) 4. E. Stötzner, T. Homburg and H. Mara, "CNN based Cuneiform Sign Detection Learned from Annotated 3D Renderings and Mapped Photographs with Illumination Augmentation," 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), Paris, France, 2023, pp. 1672-1680, doi: 10.1109/ICCVW60793.2023.00183. [↗](https://doi.org/10.1109/iccvw60793.2023.00183) 5. A. Prasetiadi and J. Saputra, “Fragmented-cuneiform-based convolutional neural network for cuneiform character recognition,” IAES International Journal of Artificial Intelligence, vol. 13, no. 1, p. 554, Mar. 2024, doi: 10.11591/ijai.v13.i1.pp554-562. [↗](https://doi.org/10.11591/ijai.v13.i1.pp554-562) 6. B. Bogacz, M. Klingmann and H. Mara, "Automating Transliteration of Cuneiform from Parallel Lines with Sparse Data," 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), Kyoto, Japan, 2017, pp. 615-620, doi: 10.1109/ICDAR.2017.106. [↗](https://doi.org/10.1109/icdar.2017.106) 7. S. Elshehaby, M. Al-Saad, A. Panthakkan and H. A. Ahmad, "Unlocking Ancient Secrets: A Deep Learning Approach to Cuneiform Symbols Recognition," 2024 Advances in Science and Engineering Technology International Conferences (ASET), Abu Dhabi, United Arab Emirates, 2024, pp. 1-6, doi: 10.1109/ASET60340.2024.10708771. [↗](https://doi.org/10.1109/aset60340.2024.10708771) 8. F. Hagelskjær, "Deep Learning Classification of Large-Scale Point Clouds: A Case Study on Cuneiform Tablets," 2022 IEEE International Conference on Image Processing (ICIP), Bordeaux, France, 2022, pp. 826-830, doi: 10.1109/ICIP46576.2022.9898032. [↗](https://doi.org/10.1109/icip46576.2022.9898032) 9. M. Mahmood, F. M. Jasem, A. A. Mukhlif, and B. Al-Khateeb, “Classifying cuneiform symbols using machine learning algorithms with unigram features on a balanced dataset,” Journal of Intelligent Systems, vol. 32, no. 1, Jan. 2023, doi: 10.1515/jisys-2023-0087. [↗](https://doi.org/10.1515/jisys-2023-0087) 10. D. Marcus, A manual of Akkadian. 1978. [Online]. Available: [http://ci.nii.ac.jp/ncid/BA03213704](http://ci.nii.ac.jp/ncid/BA03213704). [↗](http://ci.nii.ac.jp/ncid/BA03213704) 11. H. Mara and B. Bogacz, "Breaking the Code on Broken Tablets: The Learning Challenge for Annotated Cuneiform Script in Normalized 2D and 3D Datasets," 2019 International Conference on Document Analysis and Recognition (ICDAR), Sydney, NSW, Australia, 2019, pp. 148-153, doi: 10.1109/ICDAR.2019.00032. [↗](https://doi.org/10.1109/icdar.2019.00032) 12. X. Deng and C. Photong, "Evaluation of Tea Leaf Disease Identification Based on Convolutional Neural Networks VGG16, ResNet50, and DenseNet169 Image Recognitions," 2024 12th International Electrical Engineering Congress (iEECON), Pattaya, Thailand, 2024, pp. 01-04, doi: 10.1109/iEECON60677.2024.10537865. [↗](https://doi.org/10.1109/ieecon60677.2024.10537865) 13. S. Gordin et al., “Reading Akkadian cuneiform using natural language processing,” PLoS ONE, vol. 15, no. 10, p. e0240511, Oct. 2020, doi: 10.1371/journal.pone.0240511. [↗](https://doi.org/10.1371/journal.pone.0240511) 14. S. Elshehaby, M. Al-Saad, A. Panthakkan and H. A. Ahmad, "Unlocking Ancient Secrets: A Deep Learning Approach to Cuneiform Symbols Recognition," 2024 Advances in Science and Engineering Technology International Conferences (ASET), Abu Dhabi, United Arab Emirates, 2024, pp. 1-6, doi: 10.1109/ASET60340.2024.10708771. [↗](https://doi.org/10.1109/aset60340.2024.10708771) 15. Y. -n. Dong and G. -s. Liang, "Research and Discussion on Image Recognition and Classification Algorithm Based on Deep Learning," 2019 International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI), Taiyuan, China, 2019, pp. 274-278, doi: 10.1109/MLBDBI48998.2019.00061. [↗](https://doi.org/10.1109/mlbdbi48998.2019.00061) 16. S. A. Zaidi, M. S. Ansari, A. Aslam, N. Kanwal, M. Asghar, and B. Lee, “A survey of modern deep learning based object detection models,” Digital Signal Processing, vol. 126, p. 103514, Mar. 2022, doi: 10.1016/j.dsp.2022.103514. [↗](https://doi.org/10.1016/j.dsp.2022.103514) 17. G. Gutherz, S. Gordin, L. Sáenz, O. Levy, and J. Berant, “Translating Akkadian to English with neural machine translation,” PNAS Nexus, vol. 2, no. 5, May 2023, doi: 10.1093/pnasnexus/pgad096. [↗](https://doi.org/10.1093/pnasnexus/pgad096) 18. Z. Yan, K. Jin, J. Hou and P. Zhang, "An Approach to Ensemble Simplified ResNet Models-Using Egyptian Hieroglyphs Data Set," 2023 3rd International Conference on Electronic Information Engineering and Computer Science (EIECS), Changchun, China, 2023, pp. 863-871, doi: 10.1109/EIECS59936.2023.10435471. Arabic [↗](https://doi.org/10.1109/eiecs59936.2023.10435471)