This article is available at the URI as part of the NYU Library's Ancient World Digital Library in partnership with the Institute for the Study of the Ancient World (ISAW). More information about ISAW Papers is available on the ISAW website.

Except where noted, ©2014 Matteo Romanello; distributed under the terms of the Creative Commons Attribution License
Creative Commons License

ISAW Papers 7.24 (2014)

Mining Citations, Linking Texts

Matteo Romanello

Canonical citations are the standard way of citing primary sources (i.e., the ancient texts) in Classics: the ability to read them, which requires knowing what numerous abbreviations stand for, is part of the early training of any classicist. Having an expert system to capture automatically these citations and their meaning is one of the aims of the project of which the research presented in this paper is part. The desire for such a system has existed for a considerable amount of time (Crane, Seales, and Terras 2009, 26) but has yet to be solved (Romanello, Boschetti, and Crane 2009; Romanello 2013).

Such a system has great potential both for scholars in Classics and for the study of Classics as a discipline: capturing the citations of ancient texts that are contained in journal articles, commentaries, monographs and other secondary sources, allows us, for example, to track over time how and how often texts were studied, essential pieces of information for a data-driven study of the discipline and its evolution.

Another possible use of the system is to display related bibliographic references within a reading environment for ancient texts. The examples that are used in this paper are taken from work that has been done to provide the GapVis interface of the Hellespont project1 with such a functionality (see Fig. 1). One of the goals of the project is to create an enhanced virtual reading environment for one specific section of Thucydides’ Histories, the so-called “Pentecontaetia” (Thuc. 1,89 to 1,118). The references that are displayed in the secondary literature view of the reading interface are mined automatically from JSTOR and are shown together with links to the full text of the journal article as well as to the cited passage in the Perseus digital library (Romanello and Thomas 2012).

Figure 1: Secondary literature view of the GapVis-based reading interface of the Hellespont project.

Mining Citations: Extraction and Disambiguation

Extracting citations requires performing two different tasks. First, the strings that constitute the citation are captured. Second, the referent of that citation is established—the specific section of text to which the citation refers. In Natural Language Processing (NLP) jargon these two steps are called respectively Named Entity Recognition (or extraction) and Named Entity Disambiguation.

My approach to citation extraction (see Fig. 2, no. 1 and 2) is essentially based on state-of-the-art NER techniques with the only difference being what it takes to adapt these techniques to the new domain (Romanello 2013). Instead of considering only the usual named entities (NEs)–such as names of people, places and organizations–I treat as NEs the different components of a citation in addition to any mention of ancient authors and works occurring in the context that surrounds the citation itself. For this purpose four distinct entities were identified: aauthor, awork, refauwork and refscope. In its current definition, a citation is a relation between any two entities, where one is always the indication of the citation’s scope (i.e. refscope) and the other can be any of the other entities (i.e. aauthor, awork and refauwork).

Figure 2: Diagram showing the various phases of mining canonical citations from texts.

Once captured, citations need to be disambiguated: this is done by assigning to each citation its corresponding CTS URN. What this means in practice is that, for instance, the citation “Hell. 3.3.1-4” of the example showed in Fig. 2 (no. 3) is mapped to its corresponding URN, urn:cts:greekLit:tlg0032.tlg001:3.3.1-3.3.4. Designed to become the equivalent of canonical citations in a digital environment, CTS URNs are a kind of identifiers that follows the Uniform Resource Name standard and was developed within the Multitext Homer project as part of the CITE architecture to make it possible to “identify and retrieve digital representations of texts” (Smith and Blackwell 2012)2.

A Knowledge Base of Canonical Texts

NER systems of this kind typically require and rely on a surrogate of domain knowledge, such as a gazetteer or a knowledge base, to support both the extraction and disambiguation of NEs. To support the disambiguation of canonical citations such a knowledge base needs to contain, for example, all possible abbreviations of the name of an author or the title of a work, possibly in multiple languages if working on multi-lingual corpora. Since the texts we are dealing with are canonical it is possible to use this knowledge base to store, in addition to abbreviations, detailed information about the citable structures of each text such as, for example, how many books are contained in Thucydides’ Histories, how many chapters are contained in book 1 etc. Being able to query this sort of information allows one to validate the automatically extracted citations, thus making it possible to identify, if not to recover, those citations that are just impossible. An example of this phenomenon is the string “Thuc. 5. 14. 1. 41.”: although it looks as a plausible citation, it is not a valid one as the work here referred to–Thucydides’ Histories–is made of three, not four, citable, hierarchical levels, (i.e. book/chapter/section). Such errors are very common when working with OCRed texts where the lack of structural markup causes, as in this case, the footnote number to be mistakenly interpreted as being part of the canonical citation “Thuc. 5. 14. 1”.

The content in the knowledge base is structured mostly using a combination of CIDOC-CRM and FRBRoo ontologies3: the Functional Requirements for Bibliographic Records (FRBR) model, in particular, is suitable for modelling information related to Classical (canonical) texts, as was showed by Babeu et al. (2007, ref), and has influenced substantially the design of the CTS protocol. In those few cases where these ontologies did not suffice to model the data we have extended some of the classes they provide in what we called the HUmanities CITation Ontology (HuCit)4.

@prefix ecrm: <> .
@prefix efrbroo: <> .
@prefix rdfs: <> .

<> a efrbroo:F1_Work;
    ecrm:P131_is_identified_by <>;
    efrbroo:P102_has_title <>;
    owl:sameAs <> .

<> a efrbroo:F27_Work_Conception;
    efrbroo:R16_initiated <> .

<> a ecrm:E42_Identifier;
    rdfs:label "urn:cts:greekLit:tlg0003.tlg001";
    ecrm:P2_has_type <> .

<> a efrbroo:E35_Title;
    ecrm:P139_has_alternative_form <> .
    rdfs:label "Der Peloponnesische Krieg"@ger,
        "History of the Peloponnesian War"@eng,
        "La Guerra del Peloponneso"@ita,
        "l’Histoire de la guerre du Péloponnèse"@fre .  

<> a ecrm:E41_Appellation;
    rdfs:label "Thuc.";
    ecrm:P2_has_type <> .
Figure 3: The knowledge-base record for Thucydides’ Histories serialized as RDF/Turtle.

As shown in Fig. 3 our record is linked to a record in the Perseus Catalog; the CTS URN associated with the work as well as the abbreviations of its title are explicitly modelled by using respectively the CIDOC-CRM classes E42_Identifier and E41_Appellation.

Publishing Extracted Citations as Linked Open Data

Not only are canonical citations important because of their function, they are also interesting artifacts in themselves. They were designed, well before the advent of digital technologies, to refer to texts in a very precise and interoperable way: precise because texts are the fundamental object of philological research, therefore a scholarly discourse about texts needs an accurate way of referring to them; interoperable because although texts may exist in different editions and translations, scholars need to be able to refer to specific sections of them without having to worry about the many possible variations in pagination or layout each single edition may present.

If we accept that canonical citations are already a way of linking objects–i.e., the citing text and the cited text–extracting citations reconstructs and makes explicit links that already exist in the text. The act of transforming citations into hyperlinks, however, may lead to a misrepresentation of their nature and specifically of their being designed to be interoperable: a canonical citation should not be tied to the referenced passage in a specific edition, but should rather work as a resolvable pointer, that can be resolved to a given portion of text in any available edition or translation.

Let us now consider how extracted citations are stored and published online as Linked Open Data (Heath and Bizer 2011). By following an approach that was largely inspired by the Pelagios Project5, extracted canonical citations are represented as annotations as defined by the Open Annotation Data Model6 (see Fig. 4). A new annotation is created for each extracted citation: the string containing the citation becomes its label, whereas the citing and the cited texts become respectively its target and body–to use the OAC terminology–as expressed by the oac:hasTarget and oac:hasBody properties. The property oac:motivatedBy is used here to clarify the reason for creating such annotations: I chose oac:identifying as, in fact, extracting citations can be seen as the act of making explicit what is the object (i.e. text section) that is identified by a given citation.

<> a oac:Annotation;
    rdfs:label "Thuc. 1. 101";
    oac:motivatedBy oac:identifying;
    oac:hasBody <>;
    oac:hasTarget <> .
Figure 4: An extracted citation represented by means of the OAC ontology and serialized as RDF/Turtle.

The RDF fragment that is returned when the body URI is resolved (see Fig. 5) shows how the citation is not linked directly to the digital text but points to an intermediate object called hucit:TextElement7. This abstract object identifies a citable element within the hierarchical structure of a text and is linked via the hucit:resolves_to property to digital representations of the cited passage, in this case the editions and translations available in the Perseus Digital Library and via the Classical Works Knowledge Base (CWKB) resolution service. It must be pointed out, however, that linking to these resources is not, strictly and technically speaking, LOD-compliant as these URIs do not resolve (yet) to an RDF representation of the resource identified by the URI. However, as it has emerged clearly during the LAWDI event at which this paper was presented, linking resources together is the first necessary step to LOD, that it is hoped will be followed by making the underlying technology compliant with the LOD principles.

@prefix ecrm: <> .
@prefix hucit: <> .
@prefix rdfs: <> .

<> a hucit:TextElement;
    rdfs:label "book 1, chapter 101 of Thucydides' Histories"@en,
    ecrm:P1_is_identified_by [ a ecrm:E42_Identifier;
            rdfs:label "urn:cts:greekLit:tlg0003.tlg001:1.101";
            ecrm:P2_has_type <> ];
    hucit:is_part_of <>;
    hucit:precedes <>;
    hucit:precedes <>;
    hucit:resolves_to <>,
        <> .
Figure 5: RDF/Turtle representation identified by the URI


1 The Hellespont Project: Integrating Arachne and Perseus,

2 To date one of the main adopters of this technology is the Perseus project that has built on top of it to provide several functionalities of its digital library and catalog (see Almas et al. this volume).

3 The Erlangen OWL implementations of both CIDOC-CRM and FRBRoo were used: they are available respectively at and

4 The HuCit namespace is; the source code and some examples can be found in the code repository at

5 Pelagios: Enable Linked Ancient Geodata In Open Systems,

6 Open Annotation Data Model,

7 For further details about the design of HuCit see Romanello and Pasin (2013).

Works Cited

Babeu, Alison, David Bamman, Gregory Crane, Robert Kummer, and Gabriel Weaver. 2007. “Named Entity Identification and Cyberinfrastructure.” In Research and Advanced Technology for Digital Libraries, ed. László Kovács, Norbert Fuhr, and Carlo Meghini, 259–270. Springer.

Crane, Gregory, Brent Seales, and Melissa Terras. 2009. “Cyberinfrastructure for Classical Philology.” Digital Humanities Quarterly 3.

Heath, Tom, and Christian Bizer. 2011. Linked Data: Evolving the Web into a Global Data Space. Synthesis Lectures on the Semantic Web. Morgan & Claypool Publishers.

Romanello, Matteo. 2013. “Creating an Annotated Corpus for Extracting Canonical Citations from Classics-Related Texts by Using Active Annotation.” In Computational Linguistics and Intelligent Text Processing. 14th International Conference, CICLing 2013, Samos, Greece, March 24-30, 2013, Proceedings, Part I, ed. Alexander Gelbukh, 1:60–76. Springer Berlin Heidelberg. doi:10.1007/978-3-642-37247-6.

Romanello, Matteo, Federico Boschetti, and Gregory Crane. 2009. “Citations in the digital library of classics: extracting canonical references by using conditional random fields.” In Proceedings of the 2009 Workshop on Text and Citation Analysis for Scholarly Digital Libraries, 80–87. Morristown, NJ, USA: Association for Computational Linguistics.

Romanello, Matteo, and Michele Pasin. 2013. “Citations and Annotations in Classics : Old Problems and New Perspectives.” In DH-Case 2013. ACM.

Romanello, Matteo, and Agnes Thomas. 2012. “The World of Thucydides: From Texts to Artefacts and Back.” In Revive the Past. Proceeding of the 39th Conference on Computer Applications and Quantitative Methods in Archaeology. Beijing, 12-16 April 2011, ed. Mingquan Zhou, Iza Romanowska, Wu Zhongke, Xu Pengfei, and Philip Verhagen, 276–284. Amsterdam University Press.

Smith, Neel, and Christopher Blackwell. 2012. “Homer Multitext Project: documentation. An overview of the CTS URN notation.”