The multiple purposes of neural network for the recognition of the ancient Egyptian hieroglyphics complexity

This is a contribution by Andrea Barucci, Constanza Cucci, Massimiliano Franci, Marco Loschiavo and Fabrizio Argenti. For their short bios, please see here.

Artificial Intelligence (AI) and machine learning applications are spreading in every field of research, from fundamental physics to natural language processing and clinical medicine, with awesome results strongly affecting our everyday lives.

The advantages of such techniques are also numerous for Egyptian philology, at both the synchronic and diachronic levels. Some examples are the evolution of graphemic and hieroglyphic palaeography, the recognition of variants, as well as the calculation of the logographic, syllabic, and alphabetic percentage of the hieroglyphic writing system, to name just a few. The problem of ancient Egyptian language retrieval and classification has been addressed, with different purposes, in several works. 

In fact, in recent years, the relationship between artificial intelligence and the ancient Egyptian language has become ever closer. Many studies now go beyond the mere digitization of the text and deal with problems and solutions on different topics: coding of hieroglyphic signs (Buurman, Grimal, Hallof, Hainsworth & van der Plas 1988; Nederhof 2002); recognition and transliteration of signs (Rosmorduc 2008; Nederhof 2008; Franken & van Gemert 2013; Nederhof 2015); recognition of determinatives and their semantic field (Goldwasser, Harel & Nikolaev 2019); transliteration and translation of the signs inside cartouches (Duque-Domingo, Herrera, Valero & Cerrada 2017); how individual hieroglyphic signs are combined to form words (Nederhof & Rahman 2015); translation and analysis of the hieroglyphic text (Polis, Winand & Gillen 2013; Rosmorduc 2020).

This field of investigation creates interest in and of itself; it is clearly fruitful and must be examined from different points of view—ranging from Egyptology to Computer Science, Neural networking engineering, and computational analysis—given the complexity of the Egyptian language in all its facets and the usefulness of integrating the tools used for its analysis. Moreover, it is in this direction that it seems profitable to move forward by developing an integrated analytical system for every part of the problem: recognition and coding of signs, automatic text editing and recognition of the function of the sign, transliteration, and analysis of texts. The results and advantages of such an instrument are numerous (Rosmorduc 2015), both at the synchronic level and at the diachronic level, as well as at the regional level: the analysis of words, the recognition of variants, the graphemic evolutions, lexical changes, the calculation of the logographic, syllabic and alphabetic percentage of hieroglyphic writing system, to name a few.

We focused on single hieroglyph classification using the architectures known as Convolutional Neural Networks (CNN), which can be considered the best choice for visual recognition tasks. Starting from two labelled datasets of ancient Egyptian hieroglyphs, one publicly available and the other constructed by the authors, three well-known CNNs—successfully proposed for image recognition tasks—were tested. These tests were done either by using the transfer learning paradigm or by training from scratch. Inspired by the architecture of one of the earlier networks, a new CNN was also developed that was specifically tailored to the complexity of the classification problem at hand. Experimental classification tests were performed to compare the classic CNNs and the newly proposed one, referred to in the following as Glyphnet. 

Egyptian words consist of a linguistic sign with a signifier and what is signified, following the Saussurean definition. The first component represents the external aspect, merely graphic, that can be composed of one or more hieroglyphs. The second one represents the internal structure, essentially linguistics. The Egyptian hieroglyph is a complex sign composed of two elements: a semagram (or ideogram) and a phonogram. The semagram is a graphic symbol representing an idea in relation to it. A semagram can have two different values, depending on its function in the word: the proper semagram, which means the represented object indicating a word directly, and the determinative, a sign with a purely semantic and no phonetic value that functions to express the lexical field to which the word belongs. The “phonogram” may also have two different roles as well: the proper phonogram, which can indicate the phonetic value of the sign and only metaphonically the sound (or phonetic sequence), and the phonetic complement, a specific series of signs that expresses in a redundant way the sound of the sign to which they are accompanied. Given this complex nature, hieroglyphic signs provide fertile ground for the application of a deep learning approach for its recognition and classification.

Three architectures for image classification were used: a) ResNet, b) Inception-v3; c) Xception network. Beginning with them, we proposed Glyphnet, focusing on the specific task of hieroglyph recognition and tailoring the network to it. This focus led it to outperform the others in terms of performance, as well as ease of training and computational resources.

Even though we focused on the single hieroglyph classification task in the first research step, the application of deep learning techniques opened new and profitable perspectives in the field of Egyptology. In this view, the proposed work can be seen as the starting point for the implementation of much more complex goals. 

Still, there are several open issues that may benefit from the use of the proposed approaches: coding, recognition, and transliteration of hieroglyphic signs; recognition of determinatives and their semantic field; toposyntax of the hieroglyphs combined to form words; linguistic analysis of hieroglyphic texts; recognition of corrupt, rewritten, and erased signs. This could aid in moving even towards the identification of the “hand” of the scribe or the school of the sculptor.

Bibliography:

Barucci, A., Cucci, C., Franci, M., Loschiavo, M., Argenti, F. (2021) “A Deep Learning Approach to Ancient Egyptian Hieroglyphs Classification”, IEEE Access 9: 123438-123447, doi: 10.1109/ACCESS.2021.3110082.

Buurman, J., Grimal, N.-C., Hainsworth, M., Hallof, J., van der Plas, D. (1988) Inventaire des signes hiéroglyphiques en vue de leur saisie informatique: Manuel de codage des textes hiéroglyphiques en vue de leur saisie sur ordinateur. Informatique et Égyptologie 2. Mémoires de l’Académie des Inscriptions et Belle-Lettres (Nouvelle Série) 8, Paris: Institut de France.

Duque-Domingo, J., Herrera, P.J., Valero, E., Cerrada, C. (2017) “Deciphering Egyptian Hieroglyphs: Towards a New Strategy for Navigation in Museums”, Sensors 17: 589. 

Franken, M., van Gemert, J. (2013), “Automatic Egyptian hieroglyph recognition by retrieving images as texts”, in: Proceedings of the 21st ACM international conference on Multimedia: 765–768.

Goldwasser, O., Harel, H., Nikolaev, D. (2019), “Mapping the ancient Egyptian mind: Introducing iClassifier, a new platform for systematic analysis of classifiers in Egyptian and beyond”, Ancient Egypt and New Technology: The Present and Future of Computer Visualization, Virtual Reality and other Digital Humanities in Egyptology, International conference 29–30 March 2019 at Indiana University – Bloomington.

Nederhof, M.-J. (2002) “Alignment of resources on Egyptian texts based on XML”, in: Proceedings of the XIV Computer-aided Egyptology Round Table, Pisa, Italy: 1-12.

Nederhof. M.-J. (2008) “Automatic alignment of hieroglyphs and transliteration”, in: N. Strudwick (ed.), Information technology and Egyptology in 2008: Proceedings of the meeting of the Computer Working Group of the International Association of Egyptologists (Vienna, 8-11 July 2008), Piscataway, N.J.: Gorgias Press: 71-92.

Nederhof, M.-J., Rahman, F. (2017) “A probabilistic model of Ancient Egyptian writing”, in Journal of Language Modelling 5(1):131.

Polis, S., Winand, J. (2013) “The Ramses project. Methodology and practices in the annotation of Late Egyptian Texts”, in: I. Hafemann (ed.), Perspektiven einer corpusbasierten historischen Linguistik und Philologie, Internationale Tagung des Akademienvorhabens „Altägyptisches Wörter-buch“ an der Berlin-Brandenburgischen Akademie der Wissenschaften,12. – 13. Dezember 2011, Berlin: Berlin-Brandenburgische Akademie der Wissenschaften: 81-108.

Rosmorduc, S., (2008) “Automated transliteration of Egyptian hieroglyphs”, in: N. Strudwick (ed.), Information technology and Egyptology in 2008: Proceedings of the meeting of the Computer Working Group of the International Association of Egyptologists (Vienna, 8-11 July 2008), Piscataway, N.J., Gorgias Press: 167-183.

Rosmosrduc, S. (2015), “Computational Linguistics in Egyptology”, in: J. Stauder-Porchet, A. Stauder, W. Wendrich (eds.), UCLA Encyclopedia of Egyptology, Los Angeles, http://digital2.library.ucla.edu/viewItem.do?ark=21198/zz002jh4wt (accessed December 5, 2021). 

Rosmorduc, S. (2020), “Automated Transliteration of Late Egyptian Using Neural Networks: An Experiment in ‘Deep Learning’”, Lingua Aegyptia 28: 233-257.