Pelagios 3: Towards the semi-automatic annotation of toponyms in early geospatial documents

Introduction
Since place names form the underlying semantic content of almost all geographic documents, the ability to identify them in texts and images is essential in any attempt to work with, compare or interpret them. For early maps and geographic texts this ability is especially important, because while they rarely conform to standard geometries or schemas, they often provide the earliest attestations to towns, peoples, and other spatially localized phenomena. Tools, infrastructure and resources for collating, aligning, and exploiting toponyms in early maps and geographic documents would therefore have a broad and significant impact across a range of fields, including Archaeology, History, Classics, Genealogy and Modern Languages.

In this paper we showcase early work on the detection of possible toponyms in digitized texts and scanned old maps. It builds upon the successful Pelagios initiative which has been connecting a variety of heterogeneous online resources related to classical antiquity. In contrast, Pelagios 3 will extend its scope to the European, Islamic and Chinese Middle Ages, but focus predominantly on geographic works. These in turn will form a core body of material around which we hope to see a more diverse body of references accumulate in time.

Since our ultimate aim is to enable humanities scholars to annotate and discover places in documents for themselves, we discuss our use and adaption of existing open source tools within a framework that puts a premium on flexible, lightweight and easy to use resources. Moreover, that discussion will be based on two real-case scenarios, in order to demonstrate the strengths of our approach and flag up potential issues that require further attention.

Mapping places from texts: the Vicarello Goblets
Our first test case tackles the issue of extracting place names from a text. The Vicarello Goblets are a collection of four silver drinking vessels dated to around late third or early fourth century AD, engraved a land itinerary between Gades (modern Cadíz, Spain) and Rome. Each goblet indicates the road stations along the route (varying between 104 and 110 on each goblet), as well as the distances in miles between them. These unusual ‘texts’, the limited range of places that they represent, the easy identification of the majority of locations, along with the fact that there are images and transcriptions available online already, makes the Vicarello Goblets an optimal source for trialling the methodology that we will use on much larger corpora of texts, including travel guides, gazetteers, encyclopaedias and more.

To obtain annotations from the Vicarello Goblets, all the toponyms are matched against places in a URI-based gazetteer, that is, a directory of places which assigns a persistent Web address to each entity, allowing for disambiguation at a global level. The engravings on the Vicarello Goblets already represent ordered lists of toponyms. Therefore we can directly match the lists against the gazetteer based on name similarity, and then disambiguate by taking into account geographic proximity between different places in the list. For further documents which are of a less structured nature (i.e. which contain more free-form narrative text) we are experimenting with a combination of ‘geoparsing’ technologies, including the Edinburgh Geoparser and the Stanford NLP Toolkit.

Identification is only half the story, however. The data model that we have developed for Pelagios 3 allows for rich item metadata that cleanly differentiates between information about the item and information about the places that relate to it (and how). For instance, toponyms in a document may follow a certain sequence or layout. A simple mashup not only shows the toponyms from the four Vicarello Goblets on a map, but how they relate to one another as an itinerary. An information box at the bottom provides the information about the document itself (Fig. 1), while a small layer menu lets the user switch layers on and off for each individual goblet to allow immediate comparisons. Selecting a place displays a popup with a textual transcription from the Goblets, and metadata drawn from the gazetteer. What is noteworthy about this mashup, however, is not so much the map itself – for which comparable projects already exist – but rather that the map can be automatically generated from a simple Pelagios data file, containing item metadata and annotations in Open Annotation RDF format. Thus the pathway from data production to visualization is both efficient and highly scalable across large numbers of documents.

Extracting places from maps: Ptolemy’s Geographike Hyphegesis
Previous work on toponym recognition in scanned maps focuses on contemporary documents for the simple reason that old maps remain extremely difficult for machines to parse. Our proposal is to automate the identification of potential toponyms in terms of their location, extent and orientation on the map image, so that researchers can then associate the results with items in pre-existing gazetteer lists and ultimately with URI-based gazetteers. The example given here is of Ptolemy’s regional map of Ireland and Great Britain (Fig. 1), digitized by the British Library.

Our first processing phase generates a black-and-white mask image, which isolates and separates “background” from “foreground”. The next phase locates and characterises features – in our case, connected objects – on the foreground image using an algorithm that detects contours. Since toponyms often consist of multiple features, the final phase aims to connect the detected features to groups that most likely represent a single toponym. Fig. 2 shows that for our test case the algorithm detected toponyms with a high success rate, correctly locating 38 of 41 places.

Our initial work with additional (including visually more complex) maps has raised several error scenarios and prompted some initial responses:

Ornament irritation. Symbols and decorative elements that have structures in size and density (and colour) similar to toponyms frequently cause false positive detections. We expect that heuristics concerning the spatial density of matches and amount of overlap between them may be able to alleviate this problem, as these false detections exhibit distinctive clustering behaviour.
Line bleed. Toponyms that intersect or are located near lines (often borders, graticules or rhumbs), can distort the recognition result. We expect that proper tuning of image processing parameters in the first separation step (such as colour thresholds, or thresholds determining the behaviour of line removal algorithms) may be able to lower the number of such errors, but it is unlikely that they can be avoided altogether. Increasing the efficiency of human verification and correction is essential for addressing this challenge.
Toponym crosstalk. Especially in the presence of distracting elements such as lines, our algorithm can erroneously lead to toponym bounds that run across two neighbouring toponyms.. As in the case of errors caused by line bleed, it is unlikely that these can be avoided, but metrics based on the morphology of the toponym may help to detect and flag them to a human operator for verification.
Split toponyms. Our current processing approach does not specifically deal with toponyms that are split across multiple lines. An example can be found in Fig. 2, where "Alvion Insvla Britannica" is split into two separate toponyms. Once again however, morphology and the spatial proximity of features will allow us to present human operators with potential candidates for merging into single features.
Large area & curvilinear toponyms. Likewise, our heuristics are ill-suited to detect toponyms that cover large areas (e.g. regional toponyms), which are oriented significantly differently from other toponyms on the map, or which run along a curved baseline. Here, we may require a human to explicitly demarcate their bounds, although fortunately the size of such toponyms usually restricts their frequency in a given document.
While we expect that the amount of manual tuning and intervention will be further reduced by refining the processing workflow, toponym identification on old maps will never be a fully automated process. Therefore, we are also developing user interfaces and graphical tools that help both professional and public users carry out the manual work of aligning imagery to the gazetteers. On the one hand, we are experimenting with ways of re-presenting the user with re-oriented and visually enhanced image fragments so that they can be more easily interpreted. On the other we can use the spatial information in such images, and data from previous annotations, to propose likely candidates for ‘one-click’ annotations, and auto-completion of transcriptions.

Concluding Remarks
The data produced will provide us with opportunities to visualise both maps and texts in new ways. For instance, corpora of structurally similar documents, such as portolan charts, can be directly compared in terms of the places they refer to, the toponyms used, and their sequence along a coastline, in a similar manner to the itineraries described above. Alternatively, we can also blend out and replace toponyms with either modern or ancient alternatives where known, helping make these important documents easier to interpret for both scholars and public alike (Fig. 3). Most importantly we see this as the firsts steps in drawing new connections between the extraordinarily diverse range of early geospatial documents that have come down to us.

Fig. 1: Mashup showing route of the Vicarello Goblets (http://pelagios.github.io/demos/vicarello-alpha/)

Fig. 2: Part of Ptolemy’s regional map of Ireland and Great Britain. (Ca. 1480 © The British Library Board. Harley MS 7182 ff.. 60v-61.) Toponyms identified automatically and annotated with oriented bounding boxes.

Fig. 3: Original map (left) and map with original toponyms dynamically blended out and replaced with corresponding modern place names.

References
Rainer Simon, Elton Barker, Leif Isaksen. (2012). Exploring Pelagios: A Visual Browser for Geo-Tagged Datasets. In International Workshop on Supporting Users' Exploration of Digital Libraries. Eneko Agirre, Kate Fernie, Arantxa Otegi, Mark Stevenson (Eds.) Cyprus, Paphos, September 27, 2012, pp. 29 - 34.

Schmidt, M. G. (2011), A Gadibus Romam: Myth and Reality of an Ancient Route. Bulletin of the Institute of Classical Studies, 54: 71–86.

Elliott, T. & Gillies, S. (2011). Pleiades: An UnGIS for Ancient Geography. Poster presented at Digital Humanities 2011, Stanford University. Available at: http://dh2011abstracts.stanford.edu/xtf/view?docId=tei/ab-192.xml

Grover, C, Tobin, R, Byrne, K, Woollard, M, Reid, J, Dunn, S & Ball, J (2010), 'Use of the Edinburgh geoparser for georeferencing digitized historical collections' In Philosophical Transactions of the Royal Society A: Mathematical, Physical & Engineering Sciences, vol 368, no. 1925, pp. 3875-3889.

http://nlp.stanford.edu/downloads/lex-parser.shtml

See, for example, http://vici.org/

http://www.openannotation.org/

http://www.bl.uk/onlinegallery/onlineex/unvbrit/p/001hrl000007182u00060vrb.html

Rainer Simon, Peter Pilgerstorfer, Leif Isaksen, Elton Barker. (2013). Towards Semi-Automatic Annotation of Toponyms on Old Maps. In 8th International Workshop on Digital Approaches in Cartographic Heritage. Rome, Italy, September 19-20, 2013.