Documenting Italian Libraries on Wikidata: From Local Projects to a Multilayered National Knowledge Graph

(3.1.1) OpenRefine

Manual entry of individual data items is time-intensive, but the process can be greatly accelerated through semi-automated workflows when relevant data already exist digitally in structured formats.

OpenRefine (formerly Google Refine) is an open-source application designed for exploring, cleaning, and transforming heterogeneous or “messy” data.⁷ OpenRefine can be used easily even by novices to create facets to identify duplications and blocks of similar information and to structure, reconcile, and convert data between formats, as well as to enrich datasets through the integration of external sources and web-based services.

Both import operators used OpenRefine but, as experienced volunteers, they opposed the idea of large-scale imports without adequate data cleaning; consequently, the use of this tool was complemented by thorough pre- and post-import refinement.

(3.1.2) Import phases and data quality

Both initiatives followed a three-phase workflow.

1. Pre-import cleaning: Given the constraints of Wikidata’s data model and operator priorities, only part of the source information could be imported. Initial refinement was conducted in Excel, while OpenRefine—used for the first time by both operators—served primarily for the import rather than for extensive preliminary refinement.

2. Import phase: Data upload was typically preceded by small-scale test batches and subsequently carried out in segmented imports to facilitate close monitoring.

3. Post-import refinement: This phase addressed issues impractical or inefficient to resolve earlier.

Building on this framework, both initiatives underscored the import as the first step of continuous improvement towards a high-quality dataset. Such attention emerged from earlier bulk imports in the Italian community that had been executed with limited analysis of the underlying sources (Giaccai, 2021).⁸

(3.2.1) Preliminary analysis

The initiative proposed by the Tuscany region was aimed to make three homogeneous datasets about cultural institutions located in its territory available in the semantic web.

The import of the Sistema Cultura database⁹ was the first case of an Italian regional cultural database to be imported in Wikidata therefore there were no existing examples to refer to.

An in-depth analysis was conducted on the records due to the complexity of handling three distinct institution types (“library networks”,¹⁰ archives, libraries) managed by separate offices. Such records were sent as attached spreadsheet files via email on November 14, 2019 while awaiting the final update of the content by the members involved in the regional system, which happened before the end of the year (the updated files were sent again in January). The data about “library networks” was presented in a single-sheet format, while in the case of libraries and archives, data were distributed in three tabs with partially overlapping columns that were therefore recombined (see Table 1).¹¹

A meticulous effort was required to verify data quality and to decide which data to import in Wikidata in the corresponding properties. Some cells with descriptive paragraphs that would have been more suitable as Wikipedia content, were omitted. Sensitive (privacy-related) data and frequently-changing data (e.g. websites and contact emails) were initially excluded, as well as data that probably had already become obsolete (e.g. fax numbers). Sometimes data granularity was reduced in order to prepare for the import (e.g. structured street addresses were merged into a single string). Italian descriptions were created directly in the spreadsheet from the composition of different cells. Coordinates in the databases were also checked via random sampling and proved consistent.¹²

(3.2.2) Test and actual imports

All imports were concentrated between December 31, 2019 and February 28, 2020.¹³ The import was tested with the small file of “Library networks”, while “Archives” were imported on January 28 and February 13, and “Libraries” in different batches in the last week of February.

The imports were split along different values that were missing in the original file, for example in the case of libraries a subset of items with no ISIL (P791) was separately created, including 204 items originally inserted in Sistema Cultura with a placeholder ISIL code (IT-XX-[X]XXX, with an additional hyphenation after the province code).

Some imports were carried out in separate steps, spaced a few hours or days apart, as reflected in the item history—particularly in the initial phase. Shorter descriptions in languages other than Italian were added during the import.

All statements were imported with proper references including P248 (stated in). P813 (retrieved) was filled with the date when summary files were extracted and sent to the operators (see Table 1). The original files contained however the date of the first and sometimes also of the last update in the system, which were encoded as P577 (publication date) and P5017 (last update) respectively.¹⁴

(3.2.3) Refinement phase and trainings

The main refinement actions—for instance, manually verifying phone contacts or checking official websites when available—were catalogued.¹⁵ These tasks were not carried out solely by wiki volunteers; rather, in cooperation with regional authorities, a sustainable model and clear framework for ongoing data refinement and curation was established.

Responsibility for maintaining up-to-date content on Wikidata – continuing the practice of the former Sistema Cultura databases – was assigned to network centers or to individual archives, libraries, and museums. This subsequent phase included a training program for network representatives, organized in collaboration with AIB Toscana (Associazione Italiana Biblioteche Toscana, Italian Library Association – Tuscany chapter) in spring 2020. The program comprised two groups of 24 and 23 participants,¹⁶ each attending five lessons followed by two joint sessions. Each session lasted one hour and 30 minutes, totaling 18 hours of instruction. A further training session was held for three operators from the Province of Siena in October 2020.¹⁷

In the case of Tuscany the long-term goal was also to increase the coverage of the local “beni culturali” (cultural heritage assets) in open data ecosystems, thereby enhancing their visibility for educational and touristic purposes.¹⁸

The refinement phase continued thanks to the active reuse of the data and the constant monitoring by the employees of the GLAM institutions in the following years. Such reuse of data is exemplified through the integration of OpenStreetMap on the front-page map of the BiblioToscana website (Figure 1) that became public in December 2020 (Baldi et al., 2021). This system demonstrates a high degree of flexibility and maintains real-time connectivity with Wikidata, enabling continuous updates to be visualized over time. A comparative analysis of the map after nearly five years reveals numerous additions, many of which are directly attributable to the import of records from the ICCU Italian Libraries Database, as discussed in the following section.

Figure 1

(3.3.1) Preliminary analysis

At the start of the import,¹⁹ since the database contained at the time 18,883 records, part of them with minimal information, the first of many steps was to explore the data and create subsets to identify possible problems.

First and foremost, it is important to underline that only 67% (12,660) of the libraries in the database had the field “stato di registrazione” (registration status) blank; the others had a note in this field, most frequently “Biblioteca non censita” (non-registered library) or “Biblioteca non più esistente” (no longer existing library), implying that the data could be not up-to-date and thus discouraging their import.

Secondly, only 45.65% (8,621) of the records were updated to 2022 and another 12.65% (2,388) of the records were updated to 2021. 8.85% (1,671) had data older than five years, and 19.5% (3,685) lacked an update date – mostly the aforementioned non-registered or no longer existing libraries (Figure 2).

Figure 2

An in-depth analysis was conducted on the files, identifying potential problems within the JSON database. The choice was to use a less rich file, indirizzi.csv, containing “codice ISIL” (ISIL ID), “denominazione” (name of the library), “indirizzo” (address), “cap” (postal code), “città” (city), “provincia” (province), “regione” (region), “latitudine” (latitude) and “longitudine” (longitude) and adding data from the columns of the other files. The columns “latitudine” (latitude) and “longitudine” (longitude) were merged to be imported as P625 (coordinate location). Using the ISIL ID as the primary key and the add column(s) from other projects OpenRefine function, the columns “telefono” (telephone number), “email” and “url” (website url) were imported from the file territorio.csv while the columns “stato registrazione” (registration state), “anno censimento” (first census date), “data aggiornamento” (last update), “tipologia-amministrativa” (administrative type), “tipologia-funzionale” (functional type), “ente controllante” (parent organization), and “identificativo ACNP di una biblioteca” (ACNP library ID) were imported from the file biblioteche.json (See Table 2).

(3.3.2) Test and actual imports

To better comprehend other potential problems with the file thus created, a first test import was necessary. A subset of 391 rows was created filtering “tipologia-amministrativa” (administrative type): “Comune” (municipality), “tipologia-funzionale” (functional type): “Pubblica” (public), “regione” (region): Emilia Romagna and “stato registrazione” (registration state): blank.

After a time-consuming effort to clean the telephone numbers, e-mail addresses and website URLs for the subset, this information was excluded in the subsequent imports due to the overall poor quality of this data. Telephone number formatting, personal emails and Facebook page URLs used instead of institutional emails and web pages were amongst the main problems with this type of data.

The same test file of 391 public libraries of the Emilia Romagna region made apparent the problem of duplicated names of libraries: 24.04% (94) of the libraries necessitated disambiguation in the denomination. On the whole database the same problem affected 22.25% (4201) of the records. As mentioned previously, the data of the ICCU database is mostly collected through annual initiatives of data collection coordinated by ISTAT; these initiatives take the form of questionnaires sent to every library, compiled by a representative of the library (librarian, functionary, delegate or volunteer depending on the library) and collected at the regional level. Since it is common in smaller cities to have a single library, most responders simply used a generic denomination like “Biblioteca comunale”, “Biblioteca civica”, “Biblioteca diocesana”. This problem is not corrected at any stage of the data collection. Using GREL in OpenRefine the corresponding city was added to the library denomination, e.g. Biblioteca comunale → Biblioteca comunale di [Città] (Public library → Public library of [City]). The same solution was applied to libraries dedicated to the same public figure; the most frequent dedications were to Aldo Moro, Alessandro Manzoni and Cesare Pavese. In 13 cases there were actual duplicate libraries with different ISIL IDs, a report was made to ICCU that corrected the problem.

The first actual import consisted of 366 new items created and 25 existing items edited.²⁰

The import process continued preparing and uploading a subset for specialized libraries in Emilia Romagna region²¹ and then subsets for public libraries region by region. It was initially helpful to import no more than a couple of hundred items at a time to check the process for errors and inconsistencies. After the imports of Piemonte and Lombardia regions, respectively of 667²² and 1,170²³ items – the biggest of these first imports concerning public libraries – it was necessary a first major correction since every item created had labels and descriptions only in Italian. The issue was solved with two edit groups adding respectively 4,779²⁴ labels and 4,815²⁵ descriptions in English. From that moment, the imports creating new items had labels and descriptions in both languages.

Next in line were 1,130²⁶ university libraries (22 already existing and only edited) and 801²⁷ school libraries, then private libraries (850 created, 139 edited)²⁸ and finally libraries depending directly from Regions or the Italian state (811 created, 207 edited) (Figure 3).²⁹

Figure 3

A total of 11,239 library items were created.

The statements created by the imports were referenced with P248 (stated in), P791 (ISIL), P5017 (last update) with values from the column “data aggiornamento” (last update) and P813 (retrieved) for the date of the import.

(3.3.3) Refinement phase

After a long manual refinement it was possible to import a limited set of information about library dedicatees not included in the database as a separate field but extrapolated from the denomination. 616³⁰ values of P138 (named after) were added to libraries, only for the most identifiable historical figures. For the libraries dedicated to minor figures, it is necessary the effort of local librarians and historians to create their Wikidata items and disambiguate them appropriately from possible homonyms.

After an equally long refinement and reconciliation of pre-existing Wikidata items, it was possible to add 905³¹ values of P749 (parent organization).

In some cases the postal code was missing from the database as a standalone field: 1,088³² values of P281 were added after an extrapolation from the corresponding complete street address.

A more specific set of values of P31 (instance of) was added in 985³³ pre-existing library items with a generic instance of Q7075 (library).

To increase interoperability and facilitate subsequent imports and data comparison, 1,182³⁴ values of P10667 (ACNP library ID), where available in the ICCU database, were added.

The item Q113223474 (SBN hub, i.e. a subgroup of libraries participating in SBN) was created to be used as value of P31 for the 102 SBN hubs and 5,201³⁵ values of P463 (member of) added to the libraries which were members of a hub, using the list of libraries contained in the web page of each hub. The increasing number of statements can be seen in Figure 4.³⁶

Figure 4