Crossing Boundaries Between Humanities and Informatics: The Case of Egyptian Papyri

Elena L. Hertel, Stephan M. Unter, Kathrin Gabler, and Antonio Loprieno (Basel)

The corpus of papyri housed in the Museo Egizio in Turin comprises some 9000 fragments and approximately 230 larger ensembles and forms the most extensive known papyrus archive from the Pharaonic period. [1] The papyri originate from Deir el-Medina, the New Kingdom workers’ settlement on the Theban Westbank (1300–1070 BCE). In dealing with these papyri, the interdisciplinary project Crossing Boundaries: Understanding Complex Scribal Practices in Ancient Egypt, financed by the Swiss National Science Foundation (SNSF) [2] and the Fund for Scientific Research (F.R.S.–FNRS) and led by the University of Basel, the University of Liège, and the Museo Egizio in Turin, seeks to overcome the epistemological and methodological boundaries between archaeology, digital humanities, informatics, papyrology, palaeography, prosopography, and textual research. In this paper, we highlight the limitations of current virtual research environments and digital images for ancient manuscripts studies, exemplified on the papyrus fibres, and how the synergy with modern machine learning techniques can widen their usability.

In recent years, two important research axes have been developed in digital humanities: (a) the philological approach has addressed the digitalisation of large documentary corpora and developed state-of-the-art databases (Behlmer and Feder 2017); (b) the cognitive approach has focussed on the contribution that machine learning and artificial intelligence can provide to our understanding of humanities (Rosmorduc 2020). The project Crossing Boundaries: Understanding Complex Scribal Practices in Ancient Egypt applies both approaches to the study of the papyri from the Late Bronze Age village of Deir el-Medina (http://crossing-boundaries.unibas.ch/). The settlement was home to the community of workers (and their families) who built the royal tombs in the Valleys of the Kings and Queens in the Late Bronze Age (c. 1350–1000 BCE). Some 9000 papyrus fragments and 230 larger ensembles from Deir el-Medina are now housed in the Museo Egizio in Turin as a result of the acquisition in 1824 by the king of Sardinia of the antiquities collected by Bernardino Drovetti (1776–1852) during his time as French Consul in Egypt. Many of these papyri deal with the construction of the royal tombs in the Valleys of the Kings and Queens and they were therefore thought very likely to have originated in Deir el-Medina. Subsequent excavations at the beginning of the 20th century—by the Museo Egizio and Ernesto Schiaparelli (1856–1928)—brought to light more papyri. Based on joins between these papyri and those collected by Drovetti, it is possible to confirm Deir el-Medina as the provenance of the Drovetti papyri (Demarée, Gabler, Polis, forthcoming). Many of these documents are of significant historical importance, e.g. the Strike Papyrus, which records the oldest known reports on strikes in world history. Some of the manuscripts are more or less complete, but numerous tiny fragments kept in folders belong either to these ensembles or to other unidentified texts. We are especially interested in the scribal practices of the individual agents who produced the texts and in the life history of a particular category of complex documents that we have labelled heterogeneous papyri. These documents assemble texts or drawings that can be ascribed to different genres, such as accounts, poems, hymns, and letters. The papyri on which they were written were (re)used for a set period of time.

Our objectives are (a) to identify and document the fragments of papyri; (b) to join these fragments and provide a digital reconstruction of the original documents using Egyptological expertise and philological and palaeographical analysis for their transcription and translation, as well as a state-of-the-art machine learning approach, developed by a computer scientist and Egyptologist, to automatically analyse these historical documents; (c) to study the variety of texts attested on each papyrus and its material features, and reconstruct the individual scenarios that may lie behind the history of these documents, as well as generalisations concerning their history; (d) to enhance the results with data from other ancient Egyptian archives of papyri; and (e) to broaden the perspective on these papyri by comparing, both quantitatively and qualitatively, the data from Deir el-Medina with the complex scribal practices of other periods and places in ancient Egypt (Polis et al. 2020).

This approach has resulted in the development of the Turin Papyrus Online Platform (TPOP; https://collezionepapiri.museoegizio.it/), which we use as both a working and a publishing tool, and which provides free, worldwide online access to the papyri. Our data model is structured into objects—pieces of papyri which range from single fragments the size of a fingernail to pieces that are several meters long—which can be connected to (reconstructed) documents (longer manuscripts) that may in turn include several witnesses/texts. Each TPOP entry comprises one object and includes metadata about the object and its physical features, as well as features of the scripts and drawings and a full description of the text content (keywords, hieroglyphic transcription, transliteration, translations in up to four languages, etc.; Töpfer 2018). [3]

From the point of view of what we described above as the cognitive approach, we are developing a digital research tool called Virtual Light Table (VLT), which will allow scholars to access and filter the fragments in TPOP, add them to a digital worktable, and move, rotate, or flip them around to gain information and reconstruct documents. The results can be annotated, exported, saved, and exchanged among peers. In the future, we hope to add additional features such as graphical filters to improve the readability, or results derived from the machine learning research, such as algorithms that may help to find joins and position them accordingly. The VLT will be published as open-source software at the end of the project (Unter 2021; http://web.philo.ulg.ac.be/x-bound/virtual-light-table/).

The second cognitive domain within which Crossing Boundaries contributes to digital humanities is material studies. Papyrus as a writing material has been a topic of scholarly research for a long time: the first account of the production of Egyptian papyrus was given by the ancient author Pliny (Naturalis Historia 13.21–26) and the technical aspects of its production and general use are by now rather well understood (for a summary and bibliography on the topic, see e.g. Leach and Tait 2000:227–253). The standard format for ancient Egyptian manuscripts on papyrus were scrolls made by gluing several individual sheets together. The sheets were created by combining strips cut from the stem of a papyrus plant and then arranging them in horizontal and vertical rows on top of each other. This resulted in long papyrus rolls with the uppermost plant fibres running horizontally on one side and vertically on the other. To better preserve the papyrus, they were rolled up, thereby placing the side with the horizontal fibres on the inside (Parkinson and Quirke 1995:14). Due to the fact that the Egyptian script runs from right to left, the scrolls were inscribed in columns starting on the right-hand side of the roll (Figure 1).

Since the internal side of the manuscript is better protected when the papyrus is rolled up, the inside of a roll was traditionally used first by ancient scribes. The exterior side was often only used for the continuation of a text begun on the inside, or for a secondary, unrelated text. In many cases, it was left blank. The side of a papyrus sheet on which the text runs parallel to the fibres therefore normally coincides with the beginning of the text. Egyptologists generally use the same terminology as Classical Papyrologists and refer to this side as the recto and to the exterior as the verso. However, the aforementioned norm was not always observed by ancient scribes, while many incomplete documents have also been found, as well as documents that were cut up in antiquity. Eric Turner discussed the problematic aspects of the papyrological terminology in detail and pointed to the discrepancy between using the term recto to refer to the text content (to designate the side of a sheet containing the beginning of a text) or to materiality (to the side where the fibres run horizontally with respect to the text; Turner 1978, esp. 11–13). It is thus important to note that the designations front and back for papyrus sheets, as well as the definition of primary and secondary texts and any subsequent interpretations concerning the importance of individual texts or their relative chronology to each other are often difficult to discern, especially when dealing with fragmentary material.

A topic of study within the Crossing Boundaries project is the physical use and reuse of papyri in Deir el-Medina at the end of the New Kingdom. [4] One aspect of this study is to try and reconstruct the different stages in the life of a papyrus roll, sheet, or fragment. Egyptian papyri were often not only inscribed on one or both sides, but also wiped clean and used again, turning the manuscripts into palimpsests. In the case of the erased and usually illegible primary text of palimpsests, the relation between the direction of the script and the direction of fibres is often an important guide for understanding how the manuscript was used. [5] For this kind of research, close observation of material aspects of the object such as remaining ink smudges or fibre structure and direction are as important as the text written on the object.

An example of one such manuscript where material aspects reveal information about its history of use is P. Turin Cat. 2003 (Figure 2). The papyrus bears a text dated to the third year of Ramesses XI (22nd day of the first month of the Akhet season) in which a scribe called Thutmose takes account of certain commodities which appear to represent either the debt or the repayment of the debt of the water-carrier Pakharu (for information on the manuscript, text content, publications, and further reference, see http://papyri.museoegizio.it/o/98). The text is written parallel to the fibres on the side which—following convention—has been identified as the front or recto of the papyrus. The back or verso is blank except for one barely legible line of text—supposedly a docket [6] —which is written at a 90° angle in relation to the text on the recto thereby running parallel to the fibres as well. The general format of the manuscript—a rectangular shape with two sheet-joins running vertically to the text [7] —and the regular, vertical breaks in the manuscript seems to suggest that it was originally part of a larger papyrus that was kept rolled up (Eyre 2013:24). With a height of 21.5 cm and a width of 29.5 cm this papyrus conforms more or less with the standard measurements of a half size roll. [8] It is also possible that this part was cut from the roll to be used as an independent note, a hypothesis which might be argued for by pointing to the straight left and right edges of the papyrus which likely represent cuts rather than natural breaks. The presence of the apparently unrelated sign at the bottom right corner of the manuscript might suggest the opposite, however, indicating that the preserved papyrus is a fragment of a larger manuscript.

P. Turin Cat. 2003 is a palimpsest: both recto and verso show traces of ink and occasionally the faint remains of signs of a previous text across the sheet. Upon closer study, the traces of the primary inscription reveal an interesting feature: though hardly legible, a few signs preserved on the recto reveal the orientation of the primary text, which is at a 90° angle to the secondary text (which is turned 90° counter-clockwise to the orientation of the primary text). [9] On the verso, the signs of the erased text also seem to have been written at a 90° angle, but in the other direction, leaving them upside down in relation to the erased text on the recto (Figure 3).

This leads to the hypothesis that the papyrus as we see it now had a different format in its earlier life. It is possible that the manuscript held the same shape as now but was turned by 90°, though this would mean that the document had a vertical rather than the usual horizontal layout. Such a layout would be rare, but it is not unparalleled (e.g. P. Turin Cat. 2090+2096; one has to take into consideration, however, that this document is also a palimpsest that was used for several different purposes). Considering the unusual nature of such a format, one is left with the impression that this layout may not have been the first choice of the scribe but was instead motivated by the format of the available papyrus sheet. One might then speculate that P. Turin Cat. 2003 may have been reused twice, and that the last preserved text represents the third step of use (Figure 4).

It is also worth considering that the sheet of papyrus as it is preserved now could have been assembled by combining several smaller, previously used manuscripts. One such example has already been pointed out by Jaroslav Černý: P. Berlin P 10487 was made by joining two fragments, one with the horizontal and the other with the vertical fibres on top (Černý 1939:XX). It is noteworthy that the two sheet-joins of P. Turin Cat. 2003 are atypically close to each other, the first at 7 cm and the second at 21 cm. This means that, if these were the original joins of the manuscript, the sheets used to create the roll would have had a width of ca. 14 cm rather than the 23–26 cm common in the Ramesside Period (Eyre 2013:24). This might support the theory that the manuscript results from a secondary joining for the purpose of reuse, though it is certainly possible that the papyrus simply represents an exception to the standard sheet size.

Occasionally, one finds papyrus sheets and scrolls that were repaired in antiquity—for example, by gluing a strip of papyrus over a fragile section (see e.g. P. Turin Cat. 2090+2096). These kinds of secondary adjustments are signs of frequent use and reuse of the writing material and, though important for the understanding of ancient manuscript culture, often remain unnoticed. The main reason for this is most likely that details like a change in fibre structure are often hard to recognise from photographs; this is why, at the moment, a digital research environment still cannot completely substitute for work on an original.

The investigations above show that the course of the fibres or the existence and positioning of sheet joins on a papyrus can be important sources of information for the analysis and subsequent interpretation of a document. But the size and spacing of individual fibres also contain valuable information, as they represent the only reasonably constant size within an image—unlike the size of handwriting, for example—if no additional indicators such as rulers are given. The fibres are therefore a material property that it would be very desirable to be able to automatically recognise and read from an image. Herein lies a bipartite challenge: on the one hand, the recording of the objects such that sufficient structural information is contained in the digital image, and on the other hand, the development and training of algorithmic models that can automatically detect the fibres (Figure 5).

All these material properties can easily be verified and measured if there is direct access to the objects. The analyses, however, become more challenging if digitised images are the only available source. The hurdles here lie in the creation of the images: in our case, most of the material is scanned with a standard scanner designed for modern paper and written text. These devices illuminate the object with a broad light that moves along using sensors, thus removing as much shadow as possible. When scanning a structured material like papyrus, this process reduces the visual plasticity of the material, often providing a good representation of the ink but making it harder to detect the material fibres. While the fine, slightly darker lines of core strings where the fibres are slightly thicker remain visible, it is often not possible to determine with certainty which fibre orientation belongs to the front and which one to the back of the papyrus sheet. The same observation applies to photos of papyri where either the image resolution is not high enough or the lighting conditions render the papyrus flat and with little discernible fibre structure. This makes it difficult for both human scholars and any algorithmic solution to detect the uppermost fibre orientation. It would be optimal to achieve the automatic detection on these images without the requirement of additional effort and specialised equipment. We should, however, also take recording strategies that allow us to preserve more information about the three-dimensional structure of the material into consideration.

It is not feasible to perform extensive scanning procedures on the ancient papyri, as it might be necessary to do several sessions for testing purposes while work on these fragile materials should be reduced as much as possible. Consequently, we conduct our tests on modern papyrus, which is much easier to access. These sheets are usually made from either banana leaves or from the original Cyperus papyrus that was used in Egypt. From a material point of view, these substitutes should not differ too much from the ancient target case. Ideally, such a test set-up includes papyri of various manufacture and quality in order to achieve some variance within the material. It is also important to note that even modern sheets made from real papyrus differ in their colour from the ancient material which is much darker due to hydrolysis or ongoing oxidation processes (Krutzsch 2020:50).

Digital images of papyri are usually created under varying conditions, and differ according to the capturing technique (scans or photography) used, the equipment, the lighting settings (including values like intensity, colour temperature, and direction of light), and more. Naturally, this has an impact on the resulting colour values in the image and the visible material structure. We assume that minor changes to the intensity of the light and its colour temperature will not have a major impact on the fibre detection setting, while slight changes in colour tone may also occur across adjoining originals in the same lighting, depending on the state of preservation and the material conditions. Both these aspects have to be borne in mind as any algorithm used on original papyri has to account for such differences. Our primary focus, however, is the direction of the light and the resulting shadows as these allow us to infer the structure. In a first testing step, we provide a consistent lighting environment and thus exclude as many disturbances as possible from the structural features. The test images are taken using a wide range of settings, including exposure and aperture, angles of light, raking light, and transmitted light from the back. We also use different backgrounds with varying reflective properties. As papyrus is a semi-transparent material, bright backgrounds reflect more light back to the fibres while dark backgrounds absorb most of it. Preliminary tests with a scanner have shown that this reflective behaviour has some noticeable impact on the contrast and the visibility of the fibre structure in the image (Figure 6).

Another interesting problem that may not come to mind immediately when working on modern or well-preserved papyrus is induced by conservational techniques used in the past. Some of these methods interfere massively with the visual quality of the papyrus. Gauze, for example, is a net-like fabric made from silk that was applied to papyrus to provide it with additional stability. This textile usually covers the full extent of the object and thus interferes visually with the underlying natural fibres. Furthermore, these nets were often attached with a very reflective glue, which adds additional brightened regions to scans or photographs that are not determined by the papyrus itself and mask its structure. Other historic restoration methods and their remains may cause less interference, though the removal of conservation papers used to fix the fragments sometimes leads to lumps of paper fibres and glue being left behind on the papyrus surface (Figure 7). The removal of glue and backing paper might also be the reason why the ink seems to have moved on some papyri in modern times, as can be deduced from the considerable quantity of black pigments accumulating around the (modern) cracks of the manuscripts. All these and other occurrences add noise to the image which makes it a little more difficult to describe the surface structure.

As soon as enough images—in number and variety—have been taken, they must be prepared for the second step, the machine analysis. For this purpose, the papyri have to be annotated manually by assigning each individual pixel to the category it belongs to. What classification classes might be reasonable in the case of fibre detection? One possibility might be the selection of three different types: (a) pixels belonging to fibre cores, i.e. the fine, slightly darker lines that come from thicker parts of the fibres and which are often visually recognisable in an image; (b) broad, flat regions in between the fibre cores where the papyrus material is usually a little brighter and allows more light to pass through; and (c) background pixels for everything which is not relevant to fibre detection, such as the background or holes. This might also include parts of the papyrus, like writing, where the fibre structure is covered. The time-efficient creation of such annotated training material in this case remains a challenging problem and is yet to be solved.

Finally, a suitable machine learning approach must be chosen and trained on this new data. We currently envision a deep learning approach with an artificial neural network capable of semantic segmentation. Neural networks are one example of a machine learning method consisting of layered networks of so-called neurons interconnected with each other (Figure 8). Networks with many subsequent layers are called deep networks and are usually combined under the term deep learning. Some deep learning architectures have already been applied to papyrus material and show promising results, e.g. for the creation of facsimiles (Dhali et al. 2019) or the detection of text baselines (Grüning et al. 2019; Kiessling et al. 2019).

As a first step, such networks have to be trained on the material, i.e., they are confronted with a large number of images and compare their calculated results with the annotated ground truth. Here, semantic segmentation means the pixel-wise calculation of class probabilities. The machine will not return one label for the whole papyrus, but a prediction mask including class probabilities for every individual pixel. Afterwards, the network adjusts its inner connections such that, in future iterations, the result will likely be closer to the expected target; the distance between the network’s prediction and the truth is thereby iteratively minimised. After many such epochs and a large enough variation of input images, the network will hopefully adjust to the problem and make similar predictions on yet unseen images. Assuming that this method leads to useful and accurate predictions, we can then derive the average direction of fibres and their position for any similar papyrus image. [10]

With these few examples from across our research, we hope to have shown that the interaction of philology, informatics, and material studies does not only aim to enhance the knowledge of a specific textual corpus (the papyri from an Egyptian workers’ village during the Late Bronze Age) in its cultural context (a creative scribal elite with a far less rigid approach to their canonic tradition than we usually imagine), but also to embed this abundant material within the scientific discourse of the digital age in at least two ways: on the one hand, by emphasising the role played by visual simulation, in addition to written information, in our conceptual data modelling; and on the other hand, by overcoming the strict separation, inherited from our recent structuralist past, between the form and the contents of a human artefact: by making papyrus fragments fit for a virtual museum, we simultaneously make them more attractive for specialised work by researchers and to the shared emotional heart of the public.

As of 25.01.2021