Knowledge Graphs: Making Our Structured Data Stand Out

Knowledge graph visual from the Linked Open Data Cloud

26 March 2021 | Carmel McNamara, IOS Press, Amsterdam, NL
IOS Press publishes hundreds of journal articles and book chapters every month. How can we ensure all of this content stands out and is discoverable? The answer is that we make sure it is machine readable, with the data being structured in a manner that it can be interlinked with other data – making it useful through semantic queries. Using this linked data, we are developing tools for users to dive in and gain insights from all the published content. We do this through our knowledge graph.

by Carmel McNamara, IOS Press, Amsterdam, NL

IOS Press publishes hundreds of journal articles and book chapters every month. How can we ensure all of this content stands out and is discoverable? The answer is that we make sure it is machine readable, with the data being structured in a manner that it can be interlinked with other data – making it useful through semantic queries. Using this linked data, we are developing tools for users to dive in and gain insights from all the published content. We do this through our knowledge graph.

As a publisher, it is our duty to ensure that the content we publish is accessible and discoverable. Researchers publish their own articles in IOS Press journals and books – and they are also continually seeking knew knowledge and striving to extract relevant information from the preponderance of data that is out there.

Making linked, machine-readable metadata publicly available is an area in which IOS Press is invested and in line with this, in 2018, the beta version 1 of our linked data portal LD Connect was launched. The company's founder and director Einar Fredriksson commented at the time: “As a science publishing house operating in an era of digital transformation, we felt it was imperative for us to apply best practices to all aspects of our workflow. By offering datasets in machine-readable form to third parties and semantic tools, we hope to contribute in a meaningful way to scientific progress.” 

We strive to make content as accessible as possible in line with FAIR principles, which we touched upon in an earlier blog post (here). In this post, our focus is to consider how we can maximize the value of all the content we publish by ensuring users can easily extract information and carry out searches, discovering inferences and links that are relevant to them in the process.


What is a knowledge graph?

Most knowledge graphs will work on the same principle: organizing information in a structured way by explicitly describing the relations among entities [1]. Though there is still no single, universally-accepted definition of a knowledge graph, let’s go with the definition by Heiko Paulheim, published in Semantic Web in 2016: “A knowledge graph: mainly describes real world entities and their interrelations, organized in a graph; defines possible classes and relations of entities in a schema; allows for potentially interrelating arbitrary entities with each other; and covers various topical domains” (see visual below) [2].

Knowledge graph definition

A knowledge graph isn’t simply a two-axis graph, which might be the image that your mind conjures up when you hear the word “graph” – it is more like a 3D representation of a multi-dimensional network that not only informs you about fact A and fact B but also any relationship between fact A and B. For example: A may be the article [2] mentioned above and B may stand for its author Heiko Paulheim; then, the graph would carry the “is author of” relationship between B and A. When you search a knowledge graph, you can also visualize connections between datasets and often make serendipitous discoveries about linkages between components within the data. 


Navigating datasets with machine learning technology

Artificial intelligence (AI) and machine learning technology, such as deep neural networks, can help us navigate the vast body of information and process huge datasets, as long as the metadata is correctly coded of course. Communication between humans and machines is complicated for many reasons, none less so than our human language, which is full of exceptions and ambiguities that are hard for computers to learn. Therefore, clear and precise coding is of the utmost importance, with entities grouped under ontologies (see Box 1). 

Box 1

The knowledge graph is built on links between the data, which is done using “triples” in the form of subject–predicate–object expressions. Starting with an author – Jane – in one of our journal articles, one basic triple would be: Jane (subject) is a (predicate) researcher (object). Then, for example, if we look at a group of people who authored an article together, we can visualize the connections between the individuals. We can call these people our entities and the ontologies contain classes and properties. For instance, an ontology might define “person” as the class containing all individual persons. It might also define type as the property that assigns an entity to a class (“researcher,” “friend,” “relation”). Extracting facts from free text, including attributes, and then interconnecting the entries can therefore lead to links becoming clear that were not immediately evident from the raw data (see visual below).

Knowledge graph gif visualizing the links between triples of data


AI will read the text and can understand the content but will need help with:

  • disambiguation; and
  • understanding context and meaning.

This is when keywords are vital – though not just keywords. We additionally need to include a definition (for example, we would need to distinguish between the different meaning of “bright” to avoid any confusion; the system needs to know whether the content relates to intelligence or light). After all, AI only works with the information we provide and including both the keyword plus definition will ensure that there are no ambiguities (this not only applies to words but also acronyms). This is one of the reasons that a knowledge graph is not static; it is a dynamic environment. We need to ensure keywords are kept up-to-date as language evolves. 

To help with disambiguation, co-reference resolution is used to find expressions in a given text that refer to the same real-world entity. Embeddings are also used to help the machine understand the meaning of a word by placing semantically similar inputs close together in the embedding space. The more content you have, the more accurate the similarity. For instance, a search for “artificial intelligence” would also retrieve data including variations like “AI” or related terms like “machine learning” or “model based reasoning.” With the help of machine learning techniques, the data conversion pipeline keeps on improving as more data are added.

All of this will maximize the exposure of content to a wider audience as searches become more tailored, with similarities between concepts becoming self-evident. Topics that are semantically close will be centered in the same zone on the knowledge graph and connections can be visualized. You can take a look at a mammoth knowledge graph of linked open data and explore it yourself at the “The Linked Open Data Cloud” site here, which is also featured in a Data Science article by Xander Wilcke et al. from 2017 [3]. The depiction of the LOD cloud is shown in the visual below, with 1,260 datasets with 16,187 links (as of May 2020). Each vertex represents a separate dataset in the form of a knowledge graph. 


Knowledge graph from Linked Open Data Cloud


Explore our own knowledge graph with LD Connect

You can also access another working example of a knowledge graph, i.e., our own IOS Press knowledge graph, via our linked data platform LD Connect – the new and improved version of which was released at the end of 2020. This platform is being improved in a multi-phased way, with new tools being continuously developed behind the scenes (our similarity finders are in proof-of-concept in the current site and feedback is welcome!). You can read about the earlier beta version 2 launch here. Our linked datasets include, for example, metadata of journal articles and book chapters, authors, affiliations, countries, volumes, issues, series, pre-press and publication dates, (e)ISSNs, DOIs, link to full text, accessibility, keywords, pages, and abstracts – and our embeddings are based on all full text.

Pascal Hitzler, PhD, Kansas State University and co-Editor-in-Chief of Semantic Web, is enthusiastic about these developments and the fact our unsiloed data is available in this way. He comments: “Thank you for providing this data for the scientific community! Publication metadata provided as knowledge graphs using established standards – as done in LD Connect – makes it much easier for any interested party to ingest and reuse this data, combine it with other data, etc. This enables easier development of applications, such as literature search, topic tracing, reviewer finding, co-authorship network analysis, and more. I hope that other scholarly publishing houses will follow your lead.”


Our linked data platform LD Connect has recently been re-branded and re-built from scratch to make it more informative and visually appealing. You can check out the newly-designed site here and, as always, we welcome your feedback. Sign up to be kept updated with LD Connect news here!



1. “Challenges of Knowledge Graphs” by Sebastien Dery (1 December 2016), link: (last accessed: 22 February 2021).
2. “Knowledge graph refinement: A survey of approaches and evaluation methods” by Heiko Paulheim, Semantic Web, Volume 8, Issue 3, pp. 489–508 (2016), link: (last accessed: 22 February 2021).
3. “The knowledge graph as the default data model for learning on heterogeneous knowledge” by Xander Wilcke, Peter Bloem, and Victor de Boer, Data Science, Volume 1, Issue 1/2, pp. 39–57 (2017), link: (last accessed: 22 February 2021).


Further Background Information

  • Knowledge Graphs for eXplainable Artificial Intelligence: Foundations, Applications and Challenges, Eds. Ilaria Tiddi, Freddy Lécué, and Pascal Hitzler, Volume 47 of Studies on the Semantic Web (Amsterdam: IOS Press, 2020), link: (last accessed: 22 February 2021).
  • “Knowledge Graphs: Increasing Content Discoverability” by Ruth Pickering (Yewno), ConTech Live Webinar (2020), link: (last accessed: 22 February 2021).
  • For insights into how the LD Connect embeddings and Toolbox were developed by the team read: Gengchen Mai, Krzysztof Janowicz, and Bo Yan, “Combining Text Embedding and Knowledge Graph Embedding Techniques for Academic Search Engines,” in Proceedings of SemDeep-4 Workshop co-located with ISWC 2018, Oct. 8–12, 2018, Monterey, CA, USA.