In the field of digital humanities, the mode of information consumption constitutes a fundamental factor in the quality of research. The structuring of data into knowledge graphs provides a valuable tool for navigating concepts and exploring new ideas. However, the information are often spread across multiple sources with diferent data organizations (schema, taxonomy, etc.) if not, in some cases, even diferent data formats. These diferences between sources generate a fragmentation of knowledge, and, in order to obtain an efective quality consultation, they have to be explored alltogether. In this paper, an approach for exploring multiple knowledge graphs using visual representation is discussed.
In the field of digital humanities, the use of knowledge graphs to represent data, structurally and
graphically, to both address the data source and data consumption has already been demonstrated [
Another important topic in the field of content exploration through knowledge graphs, is the visual representation of the knowledge in the graphs, which is still an open field. Although there are known search tools based on semantic engines, these are often generic systems, leaving room for research on potential methodologies for specialized visualization and navigation.
A final key point consists of cross-references, i.e. contents that can be referred to each other, and implicit concepts, i.e., concepts which are taken for granted. Current Large Language Models (LLMs) proved to address those points, shifting the boundary from how concepts are expressed to the quantity of information (number of tokens required to process the source). However, the result of this extraction may need to be expanded with knowledge from other existing sources, further increasing the fragmentation and complicating the exploration.
The contribution of this article consists in a proposal for exploring multiple knowledge graphs using visual navigation.
The following of this paper is as it follows. After considering relevant work in Section 2, in Section 3, the use of knowledge graphs for digital humanities is introduced, with its relative advantages and challenges; Section 4 describes NAVIGO, the proposed framework for exploring multiple knowledge graphs; Section 5 analyzes a case study in the field of archaeology. Finally Section 6 presents concluding remarks and possible future directions.
The exploration of knowledge graphs (KGs) and graphical visualization have been the subject of numerous research projects and studies as well as the integration of multiple data sources; in the following sections, we will delve deeper into these topics, discussing the methodologies employed and the results achieved by these projects.
The GLOBDEF system [
Metaphactory [
In the theme of integrating multiple knowledge graphs, projects such as Mapping Manuscript Migrations
(MMM) [
All the projects indicated, although they carry out efective data integration, have in common the use of their own internal KG.
The goal of our research was to validate both the visual approach to knowledge graph exploration and the efectiveness of using multiple knowledge graphs simultaneously and, throughout our research journey, we have encountered a significant challenge in the form of non-uniformity among platforms that utilize a knowledge graph as their knowledge base. For instance, prominent platforms such as Google, DBpedia, and WikiData each have unique organizational structures for their data and this produces diferent navigation results, thereby creating a lack of consistency in the information retrieved from these platforms.
Diferent platforms may use diferent terminologies, taxonomies, or ontologies, further exacerbating the fragmentation issue. Consequently, integrating data from these disparate sources into coherent search results becomes a complex task and may require sophisticated mapping and alignment techniques.
In addition to the fragmentation issue, the information present in some fields is often of a general nature. This generality can be insuficient for producing precise results and it can also lead to ambiguities and uncertainties in the data, making it dificult for users to derive accurate insights from 1Cf. https://stanbol.apache.org 2Cf. https://intavia.eu the knowledge graph.
One possible solution to fragmentation is to explore data from multiple sources in real time to produce the search results. Using instead a batch integration, despite ofering various advantages in terms of performance and complexity of the system, requires the generation of a new knowledge graph within the system, with potential redundancy of information and therefore an increase in possible misalignments between sources. Because of this, we decided to focus more on real-time exploration in our research path.
Furthermore, the data must be presented in a way that ofers efective navigation and visualization to help users explore and understand the data in an intuitive and user-friendly manner.
In the process of implementing the exploration of multiple knowledge graphs, we have hypothesized two types of scenarios: • Partially related graphs (with some concepts in common) • Completely disjoint graphs
In the first scenario, it is feasible to consider one of the graphs as the base and the others as an extensions. Therefore, the navigation results are expanded with concepts from the base graph and the estension graphs. This scenario assumes that there is at least one common schema between the graphs, which can be leveraged to create a more comprehensive and interconnected knowledge base.
The second scenario, involving completely disjoint graphs, necessitates a connecting element that allows for the extension of navigation. In our current hypothesis, this element consists of an additional layer of relationships between the disjoint graphs. This layer serves as a bridge, linking concepts from diferent graphs and enabling navigation across them. This layer can be of any type, such as a translator between IRIs (e.g., a REST service) or another KG (e.g., an additional KG with the same scheme as both the graphs). By introducing this additional layer, the scenario is efectively transformed into the first case of partially related graphs.
Following this approach, it is possible to explore the contents of multiple knowledge graphs in real time, reducing the information to be processed in batches. However, it also presents challenges, such as the need to ensure the accuracy and consistency of the produced results.
During the alignment phase of multiple knowledge graphs, it is not uncommon to find two concepts with identical definitions but diferent meanings. In this case, collision management can be: • automatic; • human moderated;
Automatic management must be performed at the moment the graphs are related to each other and can use various approaches, from the simplest based on the most recent information (if the entry date is present), to one based on learning systems to recognize which colliding data is the most reliable.
Human moderated management, on the other hand, can be performed at diferent times: • if the graphs are linked with each other in a batch operation, then it is possible to introduce a moderation phase to manually validate the collisions; • if the graphs are linked in real-time, end users themselves can determine which of the sources they consider most reliable for their consultation.
Automatic management, although it significantly reduces the manual data analysis work, introduces a degree of error in the production of results, which varies for each approach, and in some cases, human moderated management might be preferable.
In this paper we will not delve into automated collision management.
A framework called NAVIGO was designed to allow the exploration of contents coming from one or more data sources in real time (therefore without the obligation of a prior batch integration) in order to both extend the quality of the results of the exploration, and to solve the problem of knowledge fragmentation. The framework can make use of both local knowledge graphs (e.g., generated by extracting concepts from non-digital data sources) and external ones (e.g., services such as Wikidata), with the possibility of extension to other possible data sources.
The framework architecture is composed of two types of components (Figure 1): • a main coordination module; • a series of management modules for the functions, further divided into – research; – content elements; – relationships between elements.
In the field of digital humanities, the field of archeology was chosen due to its peculiarity regarding data sources. In fact, in addition to having to draw on data sources of diferent nature (including, for example, paper texts and maps) which must be cross-referenced with each other, these sources are often of ancient origin, therefore not very suitable for digital conversion, and with concepts implied.
To assess the feasibility of the research endeavor, a prototype named SCIBA [
This prototype employs semantic and cartographic search methodologies. In contrast to traditional search engines that return a list of texts, documents, and metadata containing the queried keywords, SCIBA’s semantic search aims to refine and expand search outcomes.
The main data sources used in the prototype consisted of two internal sources: • Bibliography of books in digital format (with OCR); • Italian national archive of toponyms (names associated with specific geographical locations); and two external data sources: • Wikidata’s knowledge graph (Wikimedia foundation); • Wikipedia articles (Wikimedia foundation); In addition, as a secondary external source, the DBpedia knowledge graph was added.
Finally, to carry out semantic search, the native semantic engine of Wikipedia was used, as Wikidata was the base graph for all other data sources (except DBpedia).
In a batch processing phase, the internal data sources were converted into a knowledge graph extending the one of Wikidata.
The first data to be converted, using an automated system, were the bibliographic information metadata (title, author, etc.), followed by the toponyms, which required a strong initial moderation phase due to the recurring presence of abbreviations or typos. Subsequently, to associate the content of the books with the concepts represented in the Wikidata KG, the bibliography of books was analyzed using an external Semantic Text Analytics service (DandelionAPI3) capable of associating the concepts contained in the books with references to the Wikidata knowledge graph (e.g., the concept “Rome” found in the books was associated with both the Wikidata element Q220, relating to the modern city, and the element Q18287233, relating to the imperial capital). As the last phase, the books were associated with possible toponyms cited in them in an additional internal KG using a query-content matching approach
Regarding the external sources, both Wikidata and DBpedia ofer access to a KG, so no intervention was necessary.
3Cf. https://dandelion.eu • Books KG, which includes the definition of books derived from metadata; • Toponyms KG, which includes the definition of toponyms in triple format; • Book to concepts relationships KG, which links books to external concepts; • Book to toponyms relationships KG, which links books to the toponyms they cite.
When the user enters a query, starting from the results produced by a semantic reference engine (in this case Wikipedia, managed by one of the NAVIGO modules), the various KGs are queried to produce the results from which to start the exploration of the contents. For the KGs with a schema shared with Wikidata, the use of IRIs is suficient, while for DBpedia a description-based search was used.
Each source has its unique way of handling data (e.g., Wikipedia API for Wikidata to recover the article details), the data obtained from the sources are further processed by the respective modules in order to produce consultable results (e.g., the Wikidata results are used to recall the associated Wikipedia page).
The results of this workflow generate a list of search results with all possible relationships, extrapolated both from the original KG (Wikidata) and from all the additional KGs accessible from the system. Because the data sources are disjoint, users could view the source of the data and could also choose which source to enable or disable for searching.
In this case study, the major knowledge source Wikidata with the alternative option of DBpedia are independent of each other and, while beneficial in terms of diversity and coverage, this independence also introduces the potential for data collision.
During the system testing phases, we found that users of the system preferred to exercise their discretion in deciding which results to consider valid and which to discard. This observation led us to forgo the introduction of an automatic collision management system and, while this decision increased the number of ambiguous results, it also presented an opportunity to extend the exploration to a greater number of results.
During the testing phase, the validity of cross-referencing data to improve archeology search results was confirmed and the system demonstrated the validity of using multiple knowledge graphs instead of a single integrated graph to simplify integration operations. In fact, it was possible to integrate diferent data sources without the need for burdensome batch integration procedures, also allowing for the possibility of real-time choice.
The integration of multiple knowledge graphs with other data sources solves the problem of content fragmentation and and opens up potential developments in various fields. With the use of extension graphs dedicated to a specific domain, it allows for precise consultation on topics that are usually addressed in a generic way by other content search systems. This specificity can lead to more accurate and relevant results, thereby improving the eficiency and efectiveness of data-driven research.
The real-time navigation of multiple KGs ofers several advantages compared to the generation of a single integrated source, ranging from simplicity of management to the possibility of excluding sources during the search phase, as well as the elimination of batch generation processes.
However, the issue of information collision remains a significant challenge in this context. Information collision occurs when diferent data sources provide conflicting or overlapping information about the same entity or concept and, in the case of multiple knowledge graphs, each graph may also have its unique representation and interpretation of the same entity or concept. Despite the challenges it presents, information collision also opens up exciting opportunities for future research.