Structuring Resources for a Database of Balkan Criminality in the Long Nineteenth Century

Repository Location

Zenodo

https://doi.org/10.5281/zenodo.18712335

Object Name

DATABASE ON HAJDUK CRIMINALITY

Format Names and Versions

CSV, GeoJSON, txt files

Creation Dates

From 2025-03-01 to 2025-09-29

Dataset Creators

Roxana Patraș, Conceptualisation, Data Curation, Alexandru Ioan Cuza University, Iași

Constantin Răchită, Data Curation, Validation, Alexandru Ioan Cuza University, Iași

Mihai-Bogdan Atanasiu, Data Curation, Conceptualisation, Alexandru Ioan Cuza University, Iași

Ana Odochiciuc, Software, Validation, Alexandru Ioan Cuza University, Iași

Language

English, Romanian

License

The Creative Commons Attribution 4.0 International

Publication Date

2026-02-20

(3.1) Methodological Concerns

Our research underwent a digitization phase involving scanning and OCR, which, from the viewpoint of envisaged database architecture, is trivial and thus will not be dwelt upon. The docu-frame of the hajduks’ form of life (eighteenth to early twentieth century) is based on a repository of 17 CSV files (also available in GeoJSON format) that resulted from the AI-assisted, semi-automated, or manual extraction of textual and map toponyms. This step was followed by disambiguation (for instance, in cases of geolocations sharing the same toponym) and by the identification, with the assistance of OSM, of coordinates for each entry. Coordinates were geocoded using OpenStreetMap data via the Nominatim service (© OpenStreetMap contributors), licensed under ODbL v1.0.

We opted for the software QGIS, version 3.42 (https://www.qgis.org) (last accessed: 01 March 2025), a well-reviewed and user-friendly visualisation and analysis georeferencing tool, that supports various layer formats like “vector”, “raster”, “GeoJSON”, etc., among which the comma-separated value (CSV) was, with a view to broadening the database usage beyond spatial humanities, the best option. The template supported by QGIS should contain at least three basic pieces of information (Toponym/Name, Longitude, and Latitude) corresponding to each point’s coordinates (Field X and Field Y), which we duly filled in. Besides an ID column, we added diverse information about the printed source and the additional resources used in the disambiguation process. A functional column containing snippets from primary sources was necessary in the first stage of toponym extraction and checking. Then, for copyright reasons, the column “observations” was removed.

The selection of primary sources (see the list provided in Digital Humanities Laboratory, 2026b) was made with respect to imaginary, historical, social, and cultural patterns of “haiducia” as a form of life. Accordingly, the spatial criteria and toponyms thereof conjectured correspond to several research questions:

• Where do the hajduks hide? (woods/forests, inns)

• Where do the hajduks attack? (monasteries, inns, pubs, post stations, grand routes)

• Where do the hajduks hide the stolen objects? (money hoards).

Easily recognizable as forming a distinct form of life, the aforementioned patterns might be regarded as “proxies” of the hajduk phenomenon that was largely spread and quite typical for the Balkan area: criminality (crimes filed at the Principalities’ Courts of Justice or reported); nomadism and high mobility (mapped post stations and inns along the travelling routes); hiding (documented woods/forests, caves); celebration of spoils (pubs and inns); hoarding valuable goods (money, guns, clothes); fear and uncertainty (mapped urban peripheries, peri-urban areas, isolated domains). These “proxies” have been determined and then quantified by applying a toponymical reduction of space (Pricop, 2024), which, in the light of geo-criticism (Westphal, 2013), might look tedious; however, the potential complexity of the 7 resulted toponymical classes gets to the fore when they are reintegrated in the literary domain and discussed within the spheres of concepts such as “literary factuality” (Pettersson, 2020) or “poetic justice” (Nussbaum, 1995, pp. 79–123).

(3.2) Data Sourcing and Provenance

There are two categories of sources used for toponym extraction: texts and maps, the majority in printed form, which implied pre-processing tasks such as scanning, OCR, and manual clean-up of approx. 20,000 pages. The only digitized resources were the Hai-Ro Corpus, secondary data derived from it in the form of a list of georeferenced toponyms with the DARIAH geo-browser service (https://de.dariah.eu/en/geobrowser) (last accessed: 23 June 2020), and data provided by OSM-OpenStreetMap (https://www.openstreetmap.org) (last accessed: 29 September 2025). The maps have been consulted and copied at BNF, but their use as a QGIS layer required adjustments in Photoshop and mosaic-piecing: the collation of 20 map pieces and the over-colouring of roads, borders, and suchlike with a QGIS tool (Geometry Type/Line String). The textual sources include historical/archaeological studies, toponym dictionaries, geographic dictionaries, and scholarly contributions on Romanian toponymy. The visual sources include old Russian, Austrian, and Romanian maps commissioned in the first half of the nineteenth century, as well as new maps generated with GIS software (e.g. monetary findings in the region of the Principality of Moldavia). As already mentioned, the problems of difficult transliteration, faulty map localization, toponymic fluctuation and change, and various ambiguities have been addressed by adding supplementary bibliography (for instance, on gold extraction sites). To visualize the spatial distribution of our datapoints, the output of the first extraction was structured into seven toponym classes. The table below (see Table 1) specifies, for each file, the toponym class, referential frame, primary source, preprocessing tasks, source type, number of datapoints, and types of extraction that were applied.

A supplementary file containing the 63 main cities of Romania (FINAL_romania main cities.csv) was created in order to connect and overlap the two map layers in QGIS: the 1820 Russian map and, respectively, the satellite OSM standard map available through a software plugin. Both the CSV files and GeoJSON files were named to easily track their primary sources and respective acronyms/abbreviations (sigles). Some of these primary data files also served for creating derived data files, such as collated datapoints for “hajduks stays”, “hajduk memory”, “hajduk hideouts”, etc., that have been used for various experimental visualizations. Derived data have not been uploaded to our repository as they resulted from simple operations like appending and removing duplicates. The primary data are available in DATABASE ON HAJDUK CRIMINALITY (Version 2) at https://doi.org/10.5281/zenodo.18712335.

(4.1) Extraction

The extraction of toponyms was carried out manually, semi-automatically, or with the assistance of AI. Each type of primary source required customization based on the specifics of nonstructured data therein.

In the case of the Hai-Ro corpus, the toponyms were selected from a larger list of words starting with capital letters, extracted with a simple regex. This resource has not been subjected to further refinement because, in order to function as a contrast material in a factuality framework, the literary imagination that favoured the archetype of the noble rebel (the national hajduk) and the ideal of social justice (equal chances and access to resources) needs to be kept as ambiguous or as fictional as possible.

For the majority of primary sources that underwent prior scanning and OCR, the queried key terms have been determined by considering the level of structuring of the original materials (dictionary entries, indexes, abbreviations). For instance, the hajduk stays have been tracked by searching keywords such as “cârciumă” [pub], “crâșmă” [pub] and “han” [inn], but also other old and regional forms like “crâcimă” (also taking into account possible spelling errors caused by OCR recognition), as well as frequently collocated verbs like “au mas” [to stay], occurring in the documents of the Criminal Department of the Principality of Moldavia. The geographical and historical sources were queried using keywords denoting social status, such as “haiduc” [hajduk], “tâlhar” [robber, thief], and related synonyms.

Data on woods/forests from the four Romanian regions (Moldavia, Wallachia, Oltenia, and Transylvania) have been extracted with careful consideration of the various methodologies and localizations formerly employed by toponymists engaged in the research of the aforementioned historical regions. As research on Transylvanian and Dobrujan toponymy is still ongoing, we expect that more data will be added before long. In each of the four regions mentioned above, we searched for key-terms such as “pădur”, “codr”, “dendron”, “fiton” or “păd”, we identified the dictionary/thesaurus entry that contained them, and then we extracted data on surroundings (village, commune or county names) that enabled as precise georeferencing as possible.

The criminality places reported by the foreign travellers in the Romanian Principalities have been determined and checked as follows:

1. fed NotebookLM (Gemini 1.5 Pro version/accessed in May 2025) with 12 pdf-s of primary sources (the series Călători_străini);

2. prompted in a few shots the same request for all volumes in the series;

3. evaluated results provided by several NotebookLM notes and data tables;

4. manually checked information in the notes and data tables by retrieving larger contexts from pdf-s;

5. generated a list of toponyms.

The prompt used to search the material with the AI tool comprised a set of typo variants, synonyms, and keywords serving as proxies for the hajduks’ form of life (see prompts_NotebookLM.txt in Digital Humanities Laboratory, 2026a). Toponymical data extracted with AI tools represent 2.51% from the total of 6930 final places. Because rivers need a different data support (line vector and not point, currently adjoined to toponyms), river names mentioned by travellers in contexts of robbery have been “punctured” by searching for gold extraction sites along their courses.

The post stations, inns, and other types of data related to the grand routes, waters, or relief mapped by the Russians in 1820 have been transliterated from Cyrillic into the Latin alphabet. The money hoardings sites have been extracted from maps provided in Elena Arcuș-Jantovan’s contribution (2022), referring to historical regions such as Bessarabia (currently Republic of Moldova, located on the left bank of the Prut River), the Northern part of Bukovina (currently in Ukraine) and the region of Moldavia (currently Romania, located on right-bank of the Prut River).

The toponyms extracted from textual and map sources have been inserted in CSV format, with data standardised in a minimum of seven columns: A. ID; B. Name – toponyms from primary sources; C. Address – current name of an address located as close as possible to the toponym mentioned in column B (cross-checked with OSM); D. Sigle in LIST of PRIMARY SOURCES – contains a general acronym for a cluster/series of primary sources; E. Primary Source – short description of data; F. Longitude – decimal degrees, OSM; G. Latitude – decimal degrees, OSM.

Figure 1 shows the workflow that integrates all transformations from primary sources to CSV and GeoJSON files.

Figure 1

(4.2) Challenges and Limitations

The main challenge we encountered was disambiguating toponyms in primary sources. Briefly, in some cases, two or more geolocations share the same toponym, while in others a single geolocation has borne several successive names, each with its own spelling and alphabet. In both cases, we used contextual analysis, which involved manual checking of all data.

The processing of data contained by visual primary sources faced both linguistic and localization challenges because old maps of Russian or Austrian production do not have a homogeneous system of recording toponyms: for instance, on the Russian 1820 map we used, places in Transylvania that had Hungarian, Austrian, and Romanian names were inscribed, by using the Cyrillic alphabet, with only one of them. The cartographers’ preference for one of the three names lacks a clear rationale and requires methodological input from political history.

In Table 2, we provide a short list of transliteration challenges (see readme_DATABASE ON HAJDUK CRIMINALITY.txt in Digital Humanities Laboratory, 2026a):

The disambiguation of monetary-hoard data was carried out by carefully assessing place names associated with different geolocation coordinates. For instance, on the current territory of the Republic of Moldova, there are at least 3 villages named “Seliște” (in the counties Leova, Nisporeni, and Orhei). Similarly, several sites on both the left and right banks of the River Prut share the same toponym and were precisely determined following the evaluation of neighbouring areas.

The main disambiguation issue in wood/forest georeferencing stems from the fact that dendronyms are often identical to the toponyms of large estates or counties where these geographical units were located. Another challenge arises from the relatively extensive areas of some woods/forests, which create the risk of multiple coordinate sets for the same geographic referent. Therefore, we decided to differentiate dendronyms from other toponyms by conventionally adding “Pădurea” before the wood’s name. This addition did not fix the issue of shared dendronyms for different geolocations. After an alphabetic ordering, woods with the same coordinates have been removed (as duplicates), while the shared dendronyms were added an extra number in the file name (e.g. “Pădurea Dumbrava_1”, “Pădurea Dumbrava_2”, “Pădurea Dumbrava_3”, etc.). In cases of name variation, we selected the form that is the closest to the site toponym: for instance, the entry “Negreștii” in TTM contains several overlapping dendronyms “Pădurea Negrești”, “Pădurea Negreşti-Valea Mare”, “Pădurea Negreştilor”, and “Trupul Negreşti”, among which we chose “Pădurea Negrești” as the closest to the current corresponding address. In other cases, we kept the entire row of denominators: for example, the occurrences of phrases containing the noun “Codrul”/“Codrii” [forest/forests] have been preserved as such, even if this meant a relative doubling of dendronyms (e.g. “Codrul Țibăneștilor” [The Forest of Țibănești] – “Pădurea Țibăneştilor” [The Wood of Țibănești]). With contrastive data on forest raster provided by old maps and current OSM, we envisage documenting deforestation in Romanian historical regions as an extension of our repository for a criminality database.

For documents that underwent AI processing, the challenge was not disambiguation, but the risk of losing relevant data through automated search, summarization, and coding. Therefore, we decided to add supplementary historical resources in order to enable checking of data extracted by the LLM:

1. attestations about gold extraction sites on rivers;

2. various general terminology used to connote the hajduk form of life or hajduks as persons/status (“haiduc”, “hânsar”, “gheaiduc”, “haramin”, “hoț”, “tâlhar”, “bandit”, “tâlhărie”).

For instance, the old noun “gheaiduc” [hajduk] is part of a Romanian mountain toponym (Muntele Ghelarilor), while the noun “hoț” [thief] occurs in contexts of theft supply rather than as hajduk criminality. Romanian nouns “tâlhar” [robber, thief, bandit] and “tâlhărie” [robbery] as equivalents for hajduk and hajduk crime are rather scarce in our primary sources.

Our approach to data extraction has limitations that need to be addressed in the future. The first is the relative imprecision of primary resources, chiefly maps. The second derives from the lack of homogeneity in the methodologies, descriptions of geographical neighbourhoods, and relative localizations used by Romanian toponymists. The third limit is in the eye of the beholder, which means that coordinates tagged on each list of toponyms might not be exact because of the site’s large size (Bucharest is now larger than the nineteenth century city) or because of the site’s physiographic particularities (rivers, roads, mountains) that require different types of geometries (line and polygon rather than point).

(4.3) Quality Control and Filtering

After duplicates were removed and toponym spelling errors corrected, the CSV data underwent quality control and filtering through visual inspection, consisting in the successive addition of layers in QGIS and the identification of anomalies. Visualization proved to be a very effective approach for checking positional accuracy, as files could be fixed and saved while various types of analysis were underway (e.g. buffer). In this stage, which combines conceptualization on the one hand with data collection, structuring, and processing on the other, we have paid attention to over-filtering tasks that might congest the development of this database. Conversely, we envisage enriching data sources for other fields such as archaeology, ethics, psychology, politics, and the arts. At the same time, we carefully documented the lineage of our data by creating a list of primary resources “LIST of PRIMARY SOURCES. NONSTRUCTURED to STRUCTURED DATA for a DATABASE ON HAJDUK CRIMINALITY”, which is available under https://doi.org/10.5281/zenodo.18798966. Abbreviations (sigles) were added in this bibliographic list as a means of organizing primary sources and of effectively referencing them in the CSV files.