This chapter deals with the generation and processing of research data in the Bibliotheca Hertziana’s research project “Historical Spaces in Flavio Biondo’s Italia Illustrata ”, investigating the historical understanding of social space and its evolution in the Renaissance. The study of the relations between historical maps and texts aims to explore the historical understanding of space and the knowledge associated with it by taking up approaches from cognitive science and in particular cognitive linguistics. Cognitive maps depict culture-specific spatial knowledge and practices. This knowledge is represented in different ways, which change historically through different processes and practices. The epistemological focus is therefore framed by questions such as which forms of knowledge represent spatial relations, how can processes of spatial transformation be represented and analyzed, and what is the connection between culture-specific practices and cognitive representations? In order to approach this complex of questions, the project combines cognitive-semantic parameters such as toponyms, landmarks, spatial frames of reference, geometric relations, gestalt principles, and different perspectives with computational linguistic analysis methods deployed according to our ‘Common Sense Geography’1 approach.

The project design had its starting point in the introduction and description of its general objectives, as stated in the project outline on the Bibliotheca Hertziana’s website.2 The project design is based on the assumption that precise research questions can be derived from it that will lead to operationalized hypotheses, which in turn can be tested on data. In the following, the data aspect of the project is addressed, i.e., the generation, organization, indexing, storage, and publication of research data. In principle, the underlying theoretical ([art]-historical, cognitive) framework had determined which data should be obtained from the available sources and in what way. The selection of sources was guided by various parameters: availability and effort for provision as well as the quality of digital reproductions (especially readability) of the texts and map images in digital form. For our work with geographic/cartographic data, we owe valuable suggestions to the geographic research group around Werner Kuhn3 (formerly of the Universität Münster, now UCSB) and Simon Scheider4 (previously at Universität Münster, now Universiteit Utrecht), particularly for their clear reference to cognition and knowledge representation and processing. Accordingly, we developed a three-phase structure for the project:

In the analysis phase, historical texts are processed and reinterpreted using new text and map markups and corpus-specific quantitative methods. The very general categorization pattern of spatial relations in visual perception is based on the gestalt theoretic figure-ground asymmetry adapted from the founders of cognitive linguistics, Ronald W. Langacker5 and Leonard Talmy.6 This asymmetry is not only constitutive in visual perception, but also in components of linguistic meaning in a sentence; hence, rather than clausal phrases with a subject and object, the theory instead uses figure and ground. Here the “figure” is a moving or conceptually movable entity whose site, path, or orientation is conceived as a variable, whose particular value is the relevant issue. The “ground”, meanwhile, is a reference entity, one that has a stationary setting relative to a frame of reference, with respect to which the figure’s site, path, or orientation is characterized.7

Flavio Biondo’s (1392–1463) work Italia Illustrata, published posthumously in 1474, serves as the key text. Flavio Biondo is rightly regarded as the true founder of archaeological science and antiquarian topography. Particularly in his Italia Illustrata, the first premodern description of Italy, he draws on famous authors of Roman antiquity such as Livy, Virgil, and Pliny. Greek authors such as Strabo and Ptolemy, on the other hand, are rarely considered – despite their geographical and topographical content and despite the intentions of Biondo’s work. An exception is the Latium chapter (“Regio Latina”), in which Biondo draws intensively on a hitherto undiscovered Latin translation of Strabo. Strabo’s work not only serves as a data basis but also as a structural principle: Biondo uses Strabo’s hodological description technique to locate all the Latin cities and settlements relative to each other and transforms them into a narrative structure on the west coast of Italy and three Roman roads (Via Appia, Via Latina, Via Valeria). This raises the question of whether the Italia Illustrata was actually written using many maps, as current Biondo research8 generally assumes. Since only a few regional sheet maps from the early fifteenth century have been preserved and even literary references to cartographic knowledge are rare, this problem has to be put into a larger context and possible alternatives have to be discussed in an interdisciplinary approach. Our main hypothesis within a cognitive-linguistic framework is that Biondo’s narrative is based on cognitive maps or mental models enabling the reader to mentally triangulate different spatial references. In particular, it is necessary to discuss which strategies Biondo used to collect, filter, and process his heterogeneous material of historical, geographical, archaeological, and art-historical information and to translate it into a text that could be read by a contemporary audience. Because the Latium chapter (Book II) plays a key role within Italia Illustrata, the interdisciplinary project begins with the analysis and commentary of “Regio tertia: Latina”. In contrast to previous publications on Biondo, which are comprehensive in textual criticism, biographical, literary, and art-historical references as well as in the ‘afterlife’ of Biondo’s writings, we pay particular attention to the identification of toponyms and the geographical vocabulary, spatial relations, the reception of Strabo, contemporary cartography, and mental maps.

Thus, after preparing and conditioning the sources, our next step entailed elementary text statistics, concordance generation, morphosyntactic tagging, dependency parsing, and the creation of geographical annotations. Furthermore, the spatial relations expressed in the text were annotated according to the above-mentioned cognitive-linguistic parameters, especially figure–spatial_indicator–ground triples. All research data (text and map annotations) developed so far were semantically represented by an application ontology for historical texts and maps as well as annotated content and stored in the Virtual Research Environment WissKI (see section 5. Generated Research Data and Their Semantic Enhancement through Ontology-Based Data Modeling). They are thus available as linked open data (LOD) for publication on the Semantic Web. Further research data, especially the linguistic analyses, are stored in a repository in the framework of WissKI.

The second phase – cognitive-linguistic and (art-) historical interpretation of the data (i.e., the text compared to contemporary maps) – constitutes the center of the project. To answer questions about the historical understanding of geographic space and its shifting development in premodern times, the research topic is “cognitive maps” in the broadest sense, as stated above in the paragraph on Biondo. The interpretation phase is not the subject of this paper; the first results of a cognitive-linguistic analysis are available in the contribution by Berthele and Thiering9 in this volume.

Finally, it should be mentioned that the plan to reconstruct map sketches from the data (synthesis phase) was from the very beginning part of the agenda as a long-term project goal. This would require the following steps at the least:

First of all, there are some fundamental issues regarding the available text sources. The textual transmission of Flavio Biondo’s Italia Illustrata is rather complex due to different available editions. Modern editions are based on the surviving manuscripts,12 or the first printed edition (Biondo 1474) as supervised by Gaspare Biondo,13 or purportedly the “best-known, most-cited early printed edition” (Biondo (1531) 1559).14 The manuscripts, written and reworked over a long time span, are not uniform in orthography, punctuation, or style, thus making it difficult to use the text for digital and linguistic analyses. According to his own confession, Gaspare changed Flavio’s original for stylistic reasons before printing; and the Froben edition clearly deviates in many instances from the earlier ones. The editio princeps has been transmitted in two versions, which are, contrary to common belief, not identical, and there are, for instance, more than one hundred textual deviations between White’s and Castner’s texts in the Latium chapter alone. In the long run, our study will be extended to apply to the complete text. We decided to work with the White edition because it is still the closest one to the editio princeps; unfortunately, his English translation is not error-free and is inspected in this analysis purely for heuristic reasons and as a reference point for comparative studies.

As a prerequisite for the analysis and interpretation of Biondo’s text in the framework of cognitive semantics, we designed a workflow for text analysis, which contains automatic and semi-automatic processing steps. Its ultimate goal is to annotate, that is, to markup individual words or word sequences such as proper names of places (and also of persons and events) in different text sources as well as spatial relations in texts in the framework of cognitive semantics. Moreover, these annotations will be related to similar annotations in (historical) geographical maps.

Basically, we worked with two UTF-815 encoded text representations, plain text and TEI/XML.16 All texts have been generated digitally, mostly based on spelling-corrected OCR (Optical Character Recognition).

To get a first overview of the language register used in our selected text passage, we ran some simple analysis procedures on the Latin text and its English translation as a whole. As for the tools, we used some command line scripts of our own, and we applied a KWIC (Key Word in Context) concordance program17 and the TreeTagger with the Stuttgart-Tübingen Tagset (STTS).18 For word lists and statistics, concordances, n-grams, and collocations (by de facto word co-occurrences), we also found the interactive AntConc19 tool and the browser-based online Voyant tools20 very useful. All results for Latin and English acquired in batch processing mode have been posted on a project webpage21 for quick lookup. They include:22

For the annotation of Latin, there are still only a few computational resources and tools available. Besides a few scanned classical lexica, there is an online version of WordNet for Latin as a lexical resource.23 Powerful lemmatizers and morphological analyzers for Latin are Collatinus,24 freely available in web-based and standalone versions, and the recent LEMLAT 3.0.25 The Perseus Digital Library provides access to its “word study tool”, an online version of a Latin morphological analyzer.26 The well-known TreeTagger27 – a tool for annotating text word by word with part-of-speech, lemma, word class, and inflection information – is available for many language models, among them Latin and English. We therefore decided to work in parallel language mode for some steps, that is, with Latin and English, as far as appropriate tools were available, synchronized sentence by sentence. Ideally, we could synchronize glossing word by word using linguistic terminology, as is known in field linguistics research.28

Our goal in the first phase of analyzing and interpreting Biondo’s text in a cognitive-semantic framework was the identification of places by toponyms or definite descriptions, and of spatial relations between them. Hence, beyond the morphosyntactic and semantic information on words as provided by the above-mentioned tools, we also needed information about grammatical structures. Parsers are computational tools for grammatical analysis beyond the word level which provide structural data about constructions and collocations and reveal the constituent and dependency structures of sentences. In particular, the latter grammatical structures are particularly important for semantic representation.29 For Latin, a promising dependency parser has only recently become available,30 which we applied to the full text with promising results.31

Finally, we aim for annotated logical forms which express the spatial relations of different geographical objects described in the text. Because up to now no general tools for spatial role labeling have been available, we opted for a semi-automatic annotation procedure.

As already mentioned, Recogito offers several formats for exporting text and image annotation results. Annotations of toponyms, but also of persons (in our case also geographically important, because names of peoples often reflect names of regions), and named events can be exported, grosso modo, on two different levels of information:

In the latter format, in particular CSV tables, the following information is contained:

Recently, a possibility for annotating relations between annotated entities in the text has been introduced in Recogito 2. Relations can be exported as CSV tables – nodes and edges lists – ready for processing with the Gephi56 graph visualization platform. Generally, the relations can be chosen arbitrarily, but we found it more useful to have a predefined set, as is given in BRAT configurations and the constraints that come with them, which is why we chose to use BRAT instead. For now, we are focusing on the figure–spatial_indicator–ground triples.

The annotations described so far are bound to the linguistic level (i.e., directly related to the text and map image “surface”). To achieve a deeper and more generic semantic level, we pursued a transition to the methodological level of general knowledge representation.57 This allows the toponyms and other place descriptions in the cognitive-linguistic spatial role annotations – primarily figure–spatial_indicator–ground constructions – to be identified, enriched with general geographic information, and linked to a variety of (online) resources. The semantic and epistemic level in which these representations are anchored is determined by domain models, so-called ‘formal ontologies’,58 which may be regarded as the conceptual core of appropriate domain theories. Their underlying abstractions are hence integrated, providing much more content about the conceptualization of space59 and the geographic domain.60

For the conceptual framework we built upon the CIDOC CRM (the conceptual reference model of the International Council for Documentation/International Council of Museums), a fairly generic (‘reference’) ontology,61 originally defined for the cultural heritage sector, and acknowledged as an ISO62 standard 21127 since 2006.63 A decisive reason for choosing the CRM and its spatiotemporal extension CRMgeo64 was that being a standard, it opens up a wide spectrum of interoperability and linking to many web resources. Ontological enrichment with the CRM as a top conceptual model, which, in its basic design, is event-based, provides, for example, a generic ‘assignment event’ which has open positions to be filled or linked with the semantic roles, regarding agent, constituents (material and immaterial), time span, and place.65

A remarkable feature of Recogito is that – if defined by the gazetteer – the places are further tagged with controlled terms from a thesaurus, for example the Pleiades vocabulary.66 It allows for a more detailed characterization of the named place (for example, as a settlement, a river, or a mountain), which can immediately be integrated with the ontology-based representation. The same controlled vocabulary is used for classifying the linguistically annotated entities in spatial role labeling, defined in a BRAT configuration.

With formal ontologies, we provide an answer to the question: what is the meaning of annotations? And, at the same time, in particular with the use of a standardized formal ontology like CRM, we can directly transform the annotations into semantic representations ready for publication as linked open data.

First of all, we defined a domain ontology for the description of historical maps and their content (“hmap:”), connected to the generic CRM/CRMgeo (“ecrm:” for Erlangen CRM), based on an extension of an ontology we had already developed for a database of medieval maps several years ago.67 The ontology offers a framework for the general metadata of maps and geographical texts as well as for descriptions of their content as provided by the above-mentioned annotations.68 The meaning of each metadata component (property) is defined by a so-called ‘ontology path’ (i.e., a sequence of triples built from entities and properties of the ontology). As an example, in a map production event (hmap:M9_Map_Production) there is an actor, the creator, defined by:

As an illustration, we shall present a small selection of metadata components of a historical map (hmap:E28_Map) in a simplified manner:

For each map we may have several images in which depicted objects are annotated. For this reason we have an analogous data model for images (hmap:M34_Image). For each annotated place (hmap:M3_Annotated_Place is a subclass of crmgeo:SP6_Declarative_Place) where the (geographical) contents of the annotations are encoded in the columns of the CSV tables, each column is transformed into a component, with similar ontology paths defined for it. The annotated place is linked to the image by:

So, for example, the path for QUOTE_TRANSCRIPTION is:

Each annotation, represented as a line in the table, has a universally unique identifier (UUID) and refers to a visual item (E36) which represents at least an inscription (and may in some cases also consist of an image like a wall or tower, etc.), for example:

There are also further data models for image series and suites or portfolios of works, such as map collections or atlases. The triples resulting from the BRAT annotation are modeled as “Annotated Connections” with a Source (figure), a Label (spatial_indicator), and a Target (ground).

Recogito also offers an export option for annotated texts in TEI/XML format. To define the semantics of TEI annotations69 of named entities (i.e., place, person, and event names in texts), a mapping is applied to the respective TEI tags (thus defining the meaning of these tags in terms of CRM), as outlined, for example, by Christian-Emil Smith Ore and Øyvind Eide.70 Of course, in terms of Recogito annotations, there is a strict equivalence between TEI tags and CSV table entries.

In recent years, semantic technologies have become increasingly popular to represent, manage, and publish data in the humanities. Therefore, virtual research environments with semantic backends are used to build complex knowledge networks. Data is exposed as triples using RDF, and important vocabularies and thesauri are available as linked data. Ontologies like the CIDOC CRM, mentioned above, are the semantic backbone of this approach and provide interoperability and data exchange beyond pure linking.

WissKI71 is a ready-to-be-used web-based virtual research environment and publishing framework that, at its core, relies on Semantic Web technologies to represent curated knowledge. The system enables digital humanists to produce high-quality linked data without having to cope with the technical issues of the Semantic Web and ontologies, in particular those of CIDOC CRM. This is achieved by defining, on the one hand, a mapping between traditional index card or tabular styles and, on the other, graph-based linked data. The mapping may be opaque to the users and only managed by a (modeling) administrator.

By default, data may be input and displayed either as structured data via forms or as free text. Free text may be input through a graphical editor and is semantically indexed in terms of named entity recognition results, calendar date specifications, mentioned events, and also technical terms as far as appropriate authority files such as gazetteers are available.

Form input provides mechanisms for error reduction by showing autocompletion hints that are backed by available authority files. From the textual annotations, RDF triples may be generated and reused as structured data. Furthermore, the system allows the upload, derivation, display, and processing of images based on IIIF (International Image Interoperability Framework).72

From a technical perspective, WissKI is based on Drupal73 (version 8), a widely used web content management system with a big and active user and developer community. It has a modular architecture with a vast variety of third-party extensions available. Being such an extension, WissKI profits from a stable core system and also from these community contributions, providing all sorts of functionality.

The statement that “[i]n the last analysis all maps are cognitive maps” by Michael Blakemore and Brian J. Harley78 indicates an important topic of recent research in the history of cartography. Georeferencing (i.e., reference to geographical locations), is the underlying principle for organizing and presenting all kinds of information in maps. In addition to the analytic perspective, as described in the above sections, the idea of cognitive maps also provides us with a synthetic view, in the sense that we will use the data found by the analytic steps to reconstruct plausible sketch maps.

Easy access to cognitive mapping can be achieved by systematically investigating the questions of “where”, “what”, and “when”.79 For the “where” question (i.e., spatial information in a proper sense), naming is an elementary means to determine identity. It should be noted, though, that “where” and “what” questions are not readily applicable to all languages. Hence, it is not clear how universal these markers are.80

For a description of places, depending on the frame of reference, a specification of states or processes and of distance and direction must be added. An example of a process specification would be a route description, how a certain place can be reached. “What” (the objective) and “when” become important for the solution of spatial problems: a set of suitable properties must be given which are useful in finding a solution by means of cognitive mapping. In other words, first of all, we have to identify the elements which are necessary for an epistemological organization of spatial knowledge. And then, in a second step, we need to develop an analogical or depictional representation suitable for computational processing. Obviously, regions and their relative positions play a key role as do directions, orientation, and distance. We argue that perceiving and identifying these elements and refering to them in discourse is an accomplishment in abstraction which has in any case also a cognitive foundation. Hence, we are considering the assembly of maps and their description in terms of those primarily qualitative categories with (qualitative) spatial reasoning in mind.

There are several formal approaches to qualitative theories of geographic (Euclidean) space which are suitable for spatial reasoning.81 Thus, the most obvious approach is to combine spatial reasoning with logical inferences in the framework of the formal ontology, expressed in OWL-DL. Laure Vieu has elaborated a theory which is particularly well suited for our approach.82 On the basis of mereology as an axiomatized part-whole relation, she provides a formalization of topological concepts as well as geometrical concepts, in particular distance and orientation in first-order logic.

The Region Connection Calculus (RCC-8),83 developed by Anthony Cohn et al. is an elementary topological theory with regard to qualitative spatial reasoning. An interesting question here is whether or not the theoretically identified primitives (in RCC-8, for instance) are not only epistemologically but also cognitively plausible. We are aware of the fact that testing the plausibility of cognitive salience is commonly outlined based on online psycholinguistic test designs with individual participants. Our argument is that a large database of historical encodings in texts and maps does indeed reveal fundamental spatial patterns grounded in general cognitive principles of, for example, visual perception processes. On the other hand, there are theoretical considerations which point in the same direction, such as the concept of ‘primary theory’ in cognition, in particular in the geographical domain.84

For the final step of generating a cognitive map representation of Biondo’s Italia Illustrata, the salient data to be considered from linguistic analysis are:

Starting with the analysis based on spatial role labeling, we plan to transform the descriptions of spatial objects and their spatial relations extracted from the text into plausible cognitive sketch maps.85

Cognitive-linguistic annotation results in the figure–spatial_indicator–ground triples – geographically adjusted and semantically enriched – that deliver a set of cognitive parameters from which we can build spatial property graphs. With the help of these graphs, we will survey the possibilities to generate plausible cognitive sketch maps, similar to the procedures outlined by Maria Vasardani and Junchul Kim and others.86 Technically, we will fall back to the approach of ‘cellular geography’ developed by the geographer Waldo Tobler in 1979,87 which, built upon a coordinate grid, offers advantages over the irregular spatial polygons following political or other borderlines. Similarly, in IBM’s LILOG project,88 Mohammed Nadjib Khenkar developed a procedure for the object-oriented representation of depictions on the basis of a (coordinate) grid, the so-called ‘cell matrices’.89 Finally, these sketch maps will be evaluated in comparison to the above-mentioned geo-visualization and contemporary maps as well.

We built a semi-automatic environment designed to facilitate annotating and analyzing historical texts and maps with linguistic and geographical content and outlined a new methodology based on the generated data which may help scholars understand and compare cognitive maps. Biondo presents a number of different spaces in his (re)construction of Italy. Our main aim is to analyze historical constructions of spaces, but also the spatial encodings from a cognitive-semantic point of view. In Thiering/Berthele 2022, referring to Biondo’s text, we present the different forms of knowledge represented in spatial relations and spatial perception as elaborated from these representations. Particularly in his many quotations of classical authors, Biondo makes frequent reference to their naming of places, landmarks, toponyms, and spatial relations. Facing the problem that many place names and ancient places were no longer extant or at least no longer identifiable in his time, he was unable to refer to many places as a reference system. The text also refers to a number of different semiotic encodings such as other texts and maps. These intertextual traces are one of the major tasks facing scholars in tackling spatial conceptions of the Renaissance with respect to the ancient world as the guiding spatial matrix.