Googling Ancient Places

Googling Ancient Places
Isaksen, Leif, Archaeology, University of Southamp, leifuss@googlemail.com
Barker, Elton, Classical Studies, The Open University, e.t.e.barker@open.ac.uk
Kansa, Eric C., School of Information, University of California, Berkeley, ekansa@ischool.berkeley.edu
Byrne, Kate, Informatics, University of Edinburgh, k.byrne@ed.ac.uk
Overview
Our presentation about the Google Ancient Places (GAP) project will demonstrate new techniques to computationally identify places referenced in scholarly texts. We will also discuss deployment of simple Web services that use resulting place identifications to help bridge across online literary and material culture collections.

Project History
Funded through the Google Digital Humanities Award program (July 2010-June 2011), GAP <http://googleancientplaces.wordpress.com/> mines a portion of the Google Books Corpus <http://books.google.com/intl/en/googlebooks/history.html> to find books related to ancient locations identified by gazetteers of the Classical Mediterranean world.

GAP builds upon the Herodotus Encoded Space-Time Imaging Archive (HESTIA) project. HESTIA <http://www.open.ac.uk/arts/Hestia/> was a two-year collaboration (2008-2010) between The Open University and the Universities of Oxford and Birmingham, funded by the UK Arts and Humanities Research Council. Its aim was to explore new methods for visualizing relationships in Herodotus’ Histories. The project explored multiple approaches, including:

mapping the frequency of references to specific locations (both spatially and in terms of linear narrative)
manually and automatically generating maps of the network connections between places.
The project made use of Greek and English versions of the text from the Perseus Digital Library <http://www.perseus.tufts.edu/> which are marked up with the Text Encoding Initiative (TEI) XML schema, including geographical locations based on automated string-matching with the Perseus internal gazetteer and Getty Thesaurus of Geographic Names. Closer analysis revealed that many of the locations were misidentifications, however, and a relatively labor-intensive process was required to correct them.
HESTIA’s use of Perseus Digital Library resources demonstrates the growing power of open infrastructure already established in Classical studies. HESTIA also helped to demonstrate the utility of visualizing locations within a narrative. However, could the approach be automated so as to scale beyond manually processing individual texts? GAP attempts to answer this question through more sophisticated computational methods and by using additional open infrastructure, especially new Semantic gazetteers (see below) such as GeoNames <http://www.geonames.org/>, and the Pleiades Project <http://pleaides.stoa.org/>.

HESTIA’s focus lies in a seminal text, the Histories by Herodotus. While primary and secondary literary sources represent key resources for Classical Studies, Classics also draws upon diverse sources of material evidence gathered from art history, architecture and archaeology (Mahony and Bodard 2010:3-5). These different sources of evidence and their associated scholarship are often highly “siloed”. Reference to Perseus or Pleiades resources can improve their interoperability. To address this issue, GAP demonstrates how open digital humanities infrastructure together with the Google Books Corpus, can be used synergistically to bridge online literary and material culture collections. GAP uses Open Context <http://opencontext.org/> to test such services. Open Context is an open-access archaeological data publication system offering wide-ranging documentation of architecture, archaeological contexts, and objects from multiple contributors (Kansa and Kansa 2007). Open Context provides a map and timeline on its splash page to enable both providers and users to quickly identify related research. The ability to identify relevant scholarly literature relating to Open Context’s material culture collections would be a ground-breaking extension to this service.

Approach
Prior experience with Herodotus’ Histories informs GAP’s methodology. In developing the HESTIA Narrative Timeline, we learned that places referenced in narrative texts generally cluster together to maintain narrative coherency. Given a set of toponyms with multiple possible identifications, the set of identifications with the shortest overall path between them is likely to be correct. In addition, we can add weight to the importance of each toponym by the number of possible locations it could refer to. Somewhat counter-intuitively, this means that small, obscure places with unusual names are much better guides to location than well-known places with many namesakes. While a useful starting point, several additional factors complicate accurate place identification:

The approach does not work well for fragments or with arbitrary higher-level structures such as the alphabetic organization of an encyclopedia.
The author may assume that the anticipated audience will be able to contextualize by other narrative elements (such as well-known individuals) and thus mention only a single location (or even none at all).
The author may contextualize by giving a territory in which the place is located. These can confuse point-based algorithms as there is no single ‘best’ point that represents them.
The author may have confused the place they are discussing with another, especially if they are commenting on another work or reporting independent sources.
Occasionally the pattern location clustering assumption simply does not hold. This is especially the case for places that do not perform an active function in the text such as personal names derived from places of origin (e.g. ‘Herodotus of Halicarnassus’).
Fortunately, new open infrastructure, especially Semantic Gazetteers such as GeoNames and Pleiades, can improve the precision of place identification. Both GeoNames and Pleiades offer open, machine-readable data curated by dedicated communities. They provide unique HTTP URIs for each place to which multiple names (toponyms), locations (such as spatial coordinates) and categories (like ‘settlement’ or ‘region’) can be assigned. These gazetteers make it much easier to handle the problem of synonymy and allow us to assign non-ambiguous and easily resolved public identifiers to places identified in the Google Books Corpus.

Nevertheless, even with the aid of gazetteers, the difficulties outlined above make place identifications highly probabilistic and uncertain, especially in cases where we find either insufficient or conflicting evidence. Hard cases can then be handled by a variety of methods, including more sophisticated but computationally expensive procedures or by manual effort of a scholar. Computationally, there are multiple levels at which we can look for clustering, including the chapter, book, and corpus (of the author or even genre). Looking at higher levels may provide us with broader contextual clues. A further advantage of working with massive digital corpora is that they frequently provide multiple translations and editions. In such cases we can use the linear chain of places in one edition to inform the processing of another and vice versa. Finally, as we process more books the system can record additional metadata about the places as well as the books. In particular it may see that in cases of homonymy, one location is much more frequently mentioned than all the others (such as the Egyptian Alexandria, as opposed to the many other cities of that name). This can help in cases where we have no other contextual clues to draw on. Google Books metadata and comparison of multiple editions found Google Books corpus may thus help resolve ambiguous place determinations in some cases.

It is also important to remember that there are some hard limits imposed on the process and some pragmatic aspects to our goals. First, we are only able to identify those places for which we have an entry in a gazetteer. Natural Language Processing available to us will not identify places previously unknown. Secondly, we are not looking for a ‘perfect’ set of results for the simple reason that natural language is ultimately indeterminate. Continued improvement of computational methods, as well as more traditional forms of scholarship will be required.

Outcomes
Scaling and adapting text processing methods developed for HESTIA for the larger Google Books Corpus represents one of the key challenges for GAP. To help evaluate the effectiveness of our approach, we first reconciled local identifiers used by the HESTIA project with Pleiades URIs. We will report on how our algorithmic approach to place identifications compares with places manually identified in HESTIA using the same raw text of the Histories as used by HESTIA. We will then report on results of our algorithmic method on the 1828 translation of the Histories provided by Google. Finally we will discuss application of our algorithms for general use on the Google Books corpus, focusing on public domain texts with Library of Congress Headings DE-DG (Greco-Roman World; Greece; Italy).

GAP provides processing results in RDF-expressed annotations for each text. Such annotations are extremely useful to software but generally less helpful for humanities researchers who typically require a human interface. Thus, GAP also provides Web mapping tools, like those on the HESTIA and Open Context websites. These interfaces enable searches in both directions – from text to places, and from a place to the texts which reference it. To lower adoption barriers, we chose RESTful Web service (based on the Atom Syndication Format and GeoJSON) design patterns (see Blanke et al. 2009; Kansa and Bissell 2010). Such services enable other developers and digital humanists to incorporate our results into other research environments and applications.

As discussed above, the GAP project makes extensive use of existing digital humanities infrastructure, especially place gazetteers such as Pleiades. In doing so, GAP helps to demonstrate the growing maturity of digital scholarship in Classical studies. Rather than standing alone as isolated, one-off efforts, digital projects increasingly complement one-another and enable future work. In this light, we hope GAP will catalyze continued research (see Rosenzweig 2007) in the text processing methods, systems design, and semantic standards required to bridge gaps across literary and material culture collections.

References:
Barker ETE, Bouzarovski S, Pelling CBR, Isaksen L. 2010 “Mapping an ancient historian in a digital age: the Herodotus Encoded Space-Text-Image Archive (HESTIA), ” Leeds International Classical Journal, 9 1-24

Blanke, T., M. Hedges, and R. Palmer 2009 “Restful services for the e-Humanities — web services that work for the e-Humanities ecosystem, ” Digital Ecosystems and Technologies, 2009. DEST '09. 3rd IEEE International Conference on Digital Ecosystems and Technologies., 637-642

Cohen, Daniel J. and Roy Rosenzweig 2006 Digital History: A Guide to Gathering, Preserving, and Presenting the Past on the Web, Philadelphia University of Pennsylvania Press (link)

Kansa, Eric C., and Ahrash N. Bissell 2010 “Web Syndication Approaches for Sharing Primary Data in "Small Science" Domains., ” Data Science Journal, 9 42-53

Kansa, E., and S. Whitcher Kansa 2007 “Open Context: Collaborative Data Publication to Bridge Field Research and Museum Collections, ” International Cultural Heritage Informatics Meeting (ICHIM07): Proceedings [J. Trant and D. Bearman (eds)], Toronto: Archives & Museum Informatics, (link)

Mahony, Simon and Gabriel Bodard 2010 “Introduction, ” Digital Research in the Study of Classical Antiquity [Simon Mahony and Gabriel Bodard (eds.)], London Ashgate 1-14

Rosenzweig, Roy 2007 “Collaboration and the cyberinfrastructure: Academic collaboration with museums and libraries in the digital era, ” First Monday, 12(7)