1.1 OpenCitations Datasets

OpenCitations currently maintains two primary datasets. The first dataset is OpenCitations Index, the unified collection of open citations ingested from different sources created starting from the experience done with its first index, i.e. COCI (Heibi et al. 2019), that gathered citations from Crossref. The other dataset is OpenCitations Meta (Massari et al. 2024a), a collection of bibliographic metadata of all the resources included in the OpenCitations Index as citing or cited entities.

To date, the information integrated into the OpenCitations infrastructure comes from Crossref ² (Hendricks et al. 2020), DataCite³ (DataCite Metadata Working Group, 2021), PubMed (Hutchins et al. 2019; Canese & Weis, 2013), OpenAIRE (La Bruzzo et al. 2023; Alexiou et al. 2016), and – since November 2023 – the Japan Link Center (JaLC) (斉史 et al. 2012).

OpenCitations Meta data are available as CSV and RDF dumps. The stored metadata for the bibliographic resources includes document IDs, titles, authors, publication dates, venue information, volume and issue identifiers, page ranges, resource types, publishers, and editors. Instead, OpenCitations Index data is made available in CSV, N-Triples, and Scholix (JSON) formats, and it exposes the Open Citation Identifier (OCI) (Peroni & Shotton 2019) for the citation, the OpenCitations Meta Identifier (OMID, a persistent identifier for the entities included in OpenCitations Meta) of both entities involved in the citation, the citation creation date, the timespan between the publication dates of the cited and citing entities, and fields indicating whether both entities are published in the same journal and if they share at least one common author.

Both OpenCitations Index and Meta data are accompanied by provenance information (Massari et al. 2023), which includes the responsible agent, the source URL, and the creation and modification date of the record, as well as tracking the changes of the data associated with an entity.

1.2 JaLC Data as a Prototype Testing Facility for a Language-agnostic Approach

This paper focuses on the process of metadata crosswalk (Chen 2015) of multilingual data provided by JaLC to the OCDM. To this end, we implemented a particular workflow (Moretti & Heibi 2023) to produce citation and bibliographic data.

The introduction of an Anglo-Japanese dataset marks OpenCitations’ first formal effort to ingest extensively non-English sources, emphasizing the importance of handling multilingual data and promoting inclusivity and global knowledge dissemination.

We avail ourselves of this opportunity to expose a methodology that covers data acquisition, curation, and the production of citation and bibliographic data, highlighting the benefits of a language-agnostic approach to data integration. Indeed, we claim that this strategy fosters the preservation of bibliodiversity, enhances scholarly research, and facilitates knowledge access to data provided by any data source, including multilingual ones.

Coherently, the integration of the JaLC collection tests the management of multilingual datasets following a language-agnostic approach that prioritizes displaying data in the original language when available, to facilitate access and reuse for an international academic and research community.

Objects names, format names, versions, and repository location

JaLC citation data are integrated into OpenCitations Index, and the bibliographic metadata provided by JaLC is published in OpenCitations Meta. The JaLC data are included in the last dumps of OpenCitations Index – in CSV (OpenCitations 2023a), RDF (OpenCitations 2023b), and Scholix (OpenCitations 2023c) formats – and OpenCitations Meta, in CSV format (OpenCitations 2023d).

Creation dates

JaLC’s ingestion process took six months, from June 1, 2023, to the dataset production on November 29, 2023.

Dataset creators

The input data was supplied by the Japan Science and Technology Agency⁴ and then analyzed and processed using custom software components. Below is the list of students and researchers affiliated with the Research Centre for Open Scholarly Metadata⁵ and the Digital Humanities Advanced Research Center of the University of Bologna⁶ who took part in the OpenCitations’ datasets creation, together with their roles:

• Marta Soricetti: Software;
• Arianna Moretti: Data curation, Software;
• Ivan Heibi: Data curation, Software;
• Arcangelo Massari: Data curation, Software;
• Silvio Peroni: Supervision, Project administration, Conceptualization;
• Elia Rizzetto: Software.

Language

For communication and dissemination purposes, the datasets’ structure follows the OpenCitations Data Model, an English-based data model. However, since the OCDM has no restrictions concerning the language choice, metadata values are exposed in the original formulation when possible. For JaLC, the data include a combination of information in English, Japanese, and, occasionally, other languages.

License

In full compliance with FAIR principles (Wilkinson et al. 2016) and the spirit of Open Science, the citation and bibliographic data produced from JaLC are released under a CC0 waiver.⁷

Repository name

Dumps are released on OpenCitations’ Figshare page,⁸ and the related links are published at https://opencitations.net/download. In addition, the data are available through other services for programmatic access, listed on the OpenCitations’ website.⁹

Publication date

The latest versions of the dumps of OpenCitations Index and OpenCitations Meta, which include JaLC data, were published on Figshare on 11 December 2023 and 30 November 2023 respectively.

3.1 Data Ingestion in OpenCitations

Originally, the data in Meta was derived exclusively from Crossref. However, between December 2022 and July 2023, data from PubMed, DataCite, and OpenAIRE were introduced, and simultaneously the software extension activities for JaLC dataset integration commenced. To facilitate software maintenance and define a systematic approach at the time of such expansions, a structural re-engineering of the ingestion and production workflow became necessary (Figure 1).

As anticipated, a section of the methodology was formalized in a workflow (Moretti and Heibi 2023) published on the Social Sciences & Humanities Open Marketplace¹⁰ (Barbot et al. 2019; Ilijašić Veršić & Ausserhofer 2019). However, while this workflow primarily focuses on citation data production, here we provide an overview of its application in the JaLC case study by exposing it in conjunction with the procedure for producing bibliographic metadata. Indeed, OpenCitations’ data integration methodology relies on the interconnection between bibliographic and citation data creation processes.

3.1.2 Development of a Software Plug-in for Data Conversion

Although the current OpenCitations software infrastructure allows for the reuse of many general components common to the processing procedure of all the input datasets, two specific extensions had to be developed for each new data source. The first is aimed at reading the structure of the source dataset, extracting the bibliographic entities data and citations, and producing output files; the second performs the metadata crosswalk between the data model used by the source¹⁵ – in this case, JaLC – and OCDM.¹⁶ These components are developed as plugins of the OpenCitations software oc_ds_converter (Moretti et al. 2024), released under an ISC license.

Since JaLC is the DOI registration agency responsible for assigning DOIs to the citing entities, these identifiers were accepted as valid without further checks. Nevertheless, no information was provided about the registration agencies that assigned DOIs to the cited entities. Thus, to avoid redundant API checks, we adopted a two-step validation process relying on an ad-hoc data storage system.

During the first iteration of the dataset, all citing DOIs are accepted as valid and stored in memory as such. Metadata CSV tables are produced, concerning the citing entities only (Figure 2). In the second iteration, the cited identifiers are analyzed to verify their validity. During this phase, we produce metadata and citations CSV tables concerning the cited entities whose DOIs proved to be valid (Figure 3).

Figure 2 

Flowchart describing the preliminary processing of citing bibliographic entities.

Figure 3 

Flowchart describing the processing of cited bibliographic entities, their validation, and the production of metadata and citation tables.

The process leverages a system involving multiple storage solutions to prevent duplications and minimize external API calls. The validation pipeline includes the checks listed below:

1. Search for the identifier in the temporary in-memory storage containing data concerning the current chunk of data being processed. If the identifier is among these data, it can be considered valid, since it was encountered and validated while processing the current data chunk.
2. If the identifier was not found in (1), search for it in the main storage containing data concerning the whole dataset. Finding the DOI here implies its validity since it was detected and validated previously.
3. If neither (1) nor (2) was successful, search for the identifier in the OpenCitations databases, containing a mapping between each identifier ever encountered while ingesting any dataset and its assigned OMID. If the DOI is found in there, it is considered valid.
4. Use ID-schema-specific API services to check the validity of the ID if none of the previous attempts was successful.

After having processed each chunk of the dataset, the data tables for bibliographic and citation information are generated. These tables (CSV) serve as input for the two tools performing the data production tasks for Meta and Index. Simultaneously, the ID validation information in the temporary memory storage is transferred to the permanent storage, collecting all the valid identifiers encountered in the dataset up to the current moment.

3.1.6 Input Dataset Analysis and Multilingualism Information Loss Forecast

OCDM allows only a single value for each metadata field (title, authors, venue, etc.). Therefore, we prioritized metadata in the original language when both the original and English versions were provided. However, an analysis of the initial JaLC dataset revealed that, in a few instances, the declared original language is not Japanese, and linguistic information is not always provided. Adding to the complexity, the dataset permits the specification of multiple values for publishers related to the same entity. This goes beyond allowing different translations of the same publisher name; it extends to cases where a single entity may have associations with multiple distinct publishers. Therefore, to avoid attributing wrong metadata information to bibliographic entities when the linguistic information is not formally specified and when the declared language differs from Japanese or English, our approach was to prioritize the first encountered value. We assumed the first available value to be the most commonly used, thereby mitigating the risk of inaccurate metadata attribution. This choice was motivated by the need for a pragmatic solution following OCDM without introducing data inconsistencies.

We sought to assess the impact of the single-language constraint by analyzing specific metadata elements in the JaLC dataset, such as the bibliographic entity title, publication venue title, and author names. In the case of JaLC, we found that the metadata attributed with the highest forecasted loss of information compared to the input dataset (since only one language can be accepted) is the citing entities’ journal title (41.44%) (Figure 5, Table 3).

Figure 4 

Language distribution in Meta bibliographic entities, calculated on Meta dump, version 5 (https://doi.org/10.6084/m9.figshare.21747461.v5). The analysis was performed on bibliographic entities with a declared title.

Figure 5 

Bar charts illustrating the analysis of multilingualism within the input dataset, categorized by bibliographic metadata fields.

Table 3

Table showing the metadata languages in the original dataset and the linguistic information loss due to OCDM constraints. The total amount of metadata provided for a field is the sum of the number of values provided solely in one language, twice the number of values supplied in two languages, and the product between the number of values provided in more than two languages and the precise number of furnished languages. The information loss is calculated as the sum of values provided in more languages out of the total calculated. The publisher’s name field has not been included in the table since it does not necessarily concern the loss of linguistic information but might involve cases where the information loss derives from having multi-publisher values.

	1 LANGUAGE	2 LANGUAGES	3+ LANGUAGES	TOTAL VALUES PROVIDED	INFORMATION LOSS WRT. THE ORIGINAL DATASET
title citing	5,701,285	1,641,895	39(3 languages)	8,985,192	1,641,973; 18.27%
title cited	217,316	12,616	0	242,548	12,616; 5.2%
authors citing	9,892,522	4,556,812	39(3 languages)	19,006,263	4,556,890; 23,98%
authors cited	308,079	157,556	0	623,191	157,556; 25.28%
journal title citing	1,137,368	2,658,678	21,213 (20,572 3 languages; 641 4 languages)	6,519,004	2,701,745; 41.44%
journal title cited	180,515	0	0	180,515	0

3.2 Tools and Software

Three OpenCitations software tools are involved in the process:

1. oc_ds_converter (Moretti et al. 2024). This tool handles metadata crosswalks from the JaLC data model to OCDM, the identifier validation, and the production of citations and bibliographic data in CSV format, meant to serve as input for subsequent processes. This software includes two nested modules, i.e. the oc_idmanager, for the validation of persistent identifiers, and the oc_data_storage, for the storage system management. On the occasion of codebase extension for JaLC integration, the Identifier Manager module was extended to handle JIDs, the identifiers assigned by J-STAGE¹⁷ to publication venues (e.g. journals). To facilitate the reuse of the software components, we released a Python package on PyPI. The latest version is 1.0.0,¹⁸ distributed on 27 October 2023.
2. oc_meta (Massari et al. 2024b). This software curates bibliographic data, deduplicates entities, assigns OMIDs, and generates a metadata dataset related to the bibliographic entities involved in OpenCitations Index citations. As output, it produces bibliographic data in RDF and CSV formats, and a dataset for provenance and change tracking in RDF. For this software, a PyPI package was released to maximize reuse potential: the latest available version is 1.2.4,¹⁹ published on 16 February 2023.
3. index (Heibi et al. 2024). The Index software produces RDF, CSV, and SCHOLIX data formats of OMID-to-OMID citations and a corresponding collection of provenance data in CSV and RDF formats. The current version can be consulted in the “meta-index” branch,²⁰ bound to be merged into the master branch once a stable version is assessed.

Each software is released under an ISC license and hosted on a public GitHub repository, developed following a Test-Driven Development approach (Tilley 2004) and monitoring the percentage of tested code features with the Python “coverage” library.²¹ Currently, 83% of both oc_ds_converter and oc_meta software code is covered by tests. All the abovementioned repositories come with README documentation and use Poetry²² as a dependency management system to facilitate maintenance and foster workflow reproduction.

3.3 Language Agnostic Approach

Despite the current dominance of English-language content on the web, language-agnostic architectures are needed to address the challenges posed by globalization and the rising demand for multilingual web accessibility. Nonetheless, a universal solution for representing, storing, and processing multilingual data (Jain & Kysliak 2022) and, in particular, bibliographic data (Bergamin & Guerrini 2022) is still lacking. Over time, OpenCitations has developed general solutions aimed at meeting the needs of a wide range of data to be incorporated into a comprehensive and constantly expanding database capable of accepting information from diverse sources irrespective of the language in which it is generated (Cameron 1997). By not allowing the specification of multiple translations for the same value, the OCDM imposes the selection of a single language for each bibliographic metadata entry, even in cases where the data source provides more options, leading to an inevitable loss of information.

Recognizing the implications of this endeavor, we emphasize the need to strike a delicate balance between preserving linguistic elements and respecting infrastructural and data model constraints, leading us to the choice of adopting a language-agnostic approach. For this reason, we decided to store metadata in the original language only, where it is possible. This choice is the most suitable in our case since it allows the preservation of bibliodiversity over global uniformity in a dominant language.

4.1 Current Dataset Overview

The infrastructure currently comprises 1,975,552,846 unique citations and 114,621,237 bibliographic entities. Based on a computational analysis of the titles of the publications performed in January 2024 with a script²⁴ exploiting the “langdetect” Python library for language detection,²⁵ 15% of the bibliographic metadata stored in OpenCitations Meta are not in English (Figure 6), representing a 6.4% increase compared to the previous version of the dataset (Figure 4) following the ingestion of JaLC data.

4.2 OpenCitations Datasets After JaLC Dataset Ingestion

The JaLC input dataset counts 7,343,638 bibliographic entities and 416,125 citations. After processing the data, we eventually included in OpenCitations 7,333,238 bibliographic entities and 396,788 citations. Until the integration of this dataset, only bibliographic entities involved in Index citations were included in OpenCitations Meta. However, from a prior analysis of the JaLC dataset, we found that 6,908,305 entities had no citations, 329,078 entities were involved in citations to publications not identified with persistent identifiers managed by OpenCitations, and only 106,255 entities cited publications identified by DOI. For this reason, and given the small number of entities to be processed, we decided to integrate all bibliographic records in the dataset with DOIs assigned by JaLC into Meta.As can be seen in Table 4, it is not unusual for different sources to have overlapping information. However, only 1,137 of the citations by JaLC are in common with Crossref, and the majority of its contribution to OpenCitations is unique.

The most recent update of the datasets dates back to 29 November 2023. According to these data, the OpenCitations Index includes information collected from five sources and comprises 1.97 billion citations involving 72,848,673 citing entities and 71,805,805 cited entities, for a total of 89,920,081 unique entities involved in at least one citation link.

4.3 JaLC Data Retrieval From OpenCitations Datasets

JaLC data were included both in Meta and Index, available in all the formats mentioned above.

To retrieve JaLC citations from Index, there are two methods. The first approach implies retrieving the identifiers of the citations from the provenance dataset and then looking for the collected OCIs in the OpenCitations Index CSV dump. More in detail:

1. Access the OpenCitations Index on the “Download” page.²⁶
2. Download the dataset Citation data sources’ info (N-Triple): information regarding the data source collection.
3. Identify OCI subjects containing the string “joci”, standing for JaLC OpenCitations Index, to locate citations from JaLC.
4. Download the citation data dataset in CSV format.
5. Match OCIs in the CSV dataset with those from N-Triples to obtain JaLC citation information.

The alternative approach is suggested for the Semantic Web experts and implies using the OpenCitations Index SPARQL Endpoint to execute a query to retrieve subjects of triples with <http://www.w3.org/ns/prov#atLocation> property and the IRI identifying the internal OpenCitations Index collection dedicated to JaLC derived data (https://w3id.org/oc/index/joci) as the object, as shown in the following SPARQL query:

SELECT ?s WHERE {
      ?s
             <http://www.w3.org/ns/prov#atLocation>
             <https://w3id.org/oc/index/joci/>.
}

To access JaLC records from OpenCitations Meta, it is key to note that primary source information is contained in the provenance data, as per the OCDM model. Each Meta entity is linked to a snapshot (prov:Entity) via prov:specializationOf. This snapshot includes a prov:hadPrimarySource property indicating the primary source. For JaLC records, the primary source is https://api.japanlinkcenter.org/. Since the provenance data are not in a triplestore, downloading the Meta Provenance dataset is necessary to identify JaLC records.

5.1 Reuse Potential

The user base of OpenCitations data includes Funders, Resource and Research Managers, Researchers, Policy Makers, Research Organisations, and Providers.²⁷ More in detail, the OpenCitations datasets benefit scholars in developing countries and professionals outside academic institutions without access to commercial citation indexes, as well as ordinary citizens seeking open data, open science partners, academic publishers, tool developers, bibliometricians, and librarians (Peroni & Shotton 2022a; Peroni & Shotton 2022b). OpenCitations’ citation collection is currently used in several projects, e.g. B!son,²⁸ Optimeta,²⁹ repositories, e.g. the Staatsbibliothek zu Berlin³⁰ and ORBi ULiege,³¹ and search tools, e.g. PURE suggest.³²

JaLC data integration opens up interesting usage prospects, especially for bibliometric studies on bibliographic metadata and citations among non-English resources. This aspect is especially relevant for bibliodiversity preservation in cultural heritage and humanities-oriented studies, in which awareness toward reproducibility and replicability of research is still being established (Peels & Bouter 2018), and the introduction of open access tools that facilitate open science practices is a necessary starting point. In addition, the adopted ingestion workflow not only paves the way for addressing the intricacies of multilingual data ingestion but also opens doors for broader applications. Researchers and institutions dealing with non-English-based data sources can leverage such workflow, adapting and customizing it to suit their specific needs.

5.2 Future Developments

As a future development, since OCDM excludes multilingual storage but the language of cataloging can differ from the users’ language preferences, data reuse limitations could be contrasted with software-driven data retrieval solutions for extemporaneous translation. In this perspective, scalable solutions to the lack of common practices for multilingual access include promoting the use of controlled terms and established value vocabularies for simplicity and cost-effectiveness (Riva 2022). However, the current state-of-the-art open-source solutions do not guarantee the level of accuracy we aim to achieve. Thus, since sophisticated tools are needed for cross-lingual information retrieval, we plan to develop software exploiting open-source technologies for data exposure only, specifically trained on translation tasks regarding bibliographic metadata. Of course, the tool would come with a clear statement of the nature of the data displayed, i.e., whether it is presented in its original form or has been translated. Recently, similar tasks were addressed with approaches to multilingual information retrieval based on using pre-trained multilingual language models (Hu et al. 2023). Such a solution would limit the storage issues caused by the maintenance of the English translations of the data, in addition to respecting the OCDM structure and maximizing the reuse potential of exposed data, therefore leveraging a balance between the need for information preservation and the constraints of physical space and data model nature.