Libraries – whether physical or online – have been protecting the privacy of their patrons for quite some time; for instance by the American Library Association (ALA) Code of Ethics in 1938 (Witt, 2017). This position is shared among the International Federation of Library Associations and Institutions (2016), the American Library Association (2014) and several other national library associations. Privacy in libraries is associated with protection from unwanted government attention (Jaeger et al., 2004), but also from commercial organisations (Corrado, 2007; Maceli, 2018).

Recommender systems are used to provide suggestions about items that are valuable to a person. While there are several techniques for building recommender systems, most are based on the same principle: create a profile of the user and her peers, extend this as much as possible and update it over time. This enables the system to know the preferences of the user and thus predict other items (Pazzani & Billsus, 2007; Ricci et al., 2011; Schafer et al., 1999). Linden et al. (2003) and Smith and Linden (2017) discuss their experiences at Amazon, spanning two decades.

For those who do not feel comfortable with the lack of privacy in connection to these type of systems, Jeckmans et al. (2013) have listed countermeasures. These include raising awareness about privacy issues and invoking specific laws dealing with personal information. As it might take quite some time before this will take effect, the authors also describe technical measures such as anonymisation, randomisation and the use of cryptography.

Instead of recommending titles based on personal data, here the contents of the titles will be used. The texts of the books and chapters are analysed using ngrams.

Ngrams are based on the relationships between words, either by examining which words tend to follow others immediately, or by looking at words that co-occur within the same documents. Two consecutive words are called “bigrams”, three consecutive words are called “trigrams”. Naturally, the number of trigrams in a text is lower compared to bigrams, while the trigrams are more specific. As we are examining a large text corpus – the text of almost 11,000 books and chapters – the total number of possible trigrams is still large.

Ngrams are used in different types of research. One application is document clustering: creating related groups of documents. Each document is represented by a numerical value. The k-means algorithm is typically used to calculate the distance between the documents and a ‘cluster means’; the goal is to all documents in clusters with the smallest numerical distance (Miao et al., 2005). Furthermore, the authors looked at the performance of several types of ngrams – ranging from bigrams to 5-grams – used in document clustering. They conclude that trigrams are roughly as accurate as 4-grams and 5-grams but are more economical in their resource usage.

Apart from clustering documents, ngrams are also deployed for author attribution. This technique aims to find the characteristics of a writer’s style and use that to define whether a certain text is written by that author. Here, the ngrams are not based on clusters of words, but clusters of characters (Kešelj et al., 2003). Eve (2019) is critical of the application of this technique to identify authors and uses it to distinguish literary genres instead.

The best-known use of ngrams is probably the Google Books Ngram Viewer. This vast corpus of books is used for cultural research. The most cited example is written by Michel et al. (2011). In this paper, the authors examine the change of language over time, but also cultural changes: the rise and demise of the celebrity of certain persons and suppression of ideas over time. This is far from the only paper based on the Google Books Ngram Viewer: a recent search on this subject in the Google Scholar search engine resulted in over 3,700 titles3.

The experiment of this paper does not quite fit within these three research types. It is clearly not meant to discover long term trends, in the manner of the Google Books research. Finding authors is also not necessary: this information is provided in the metadata of the OAPEN Library. The k-means algorithm is a more general-purpose application, aimed to be useful in various situations.

The most closely connected experiments are those aiming to extract keywords from an article or book (Rohini & Ambati, 2007; Souza & Raghavan, 2014). The authors describe the use of statistical methods to find distinctive words. However, the text corpora used are small and no attempt is made to connect multiple titles.

The algorithm used in this experiment is optimised for a very specific purpose: instead of creating amorph groups, it aims to find exact relations for each individual title. These relations – based on the number of shared trigrams – are ranked. The ranking and the number of shared trigrams can be used to create services for digital libraries with an open access collection. This algorithm is not general-purpose but optimised for one specific environment.

Several other experiments have been conducted on the OAPEN Library collection: creating groups of books based on usage data (Snijder, 2019) or categorising titles based on Wikipedia pages (Snijder, 2021). In the first experiment, the download patterns are analysed to find which books are regularly selected together. So, instead of looking at individual preferences, social network analysis was deployed to find the preferences of groups of people. The more recent investigation aimed to categorize books by automatically finding the Wikipedia pages that describe their contents.

Grouping books based on usage data has drawbacks: apart from the reliance on external usage data, the results need to be interpreted. The interpretation depends on analysing aspects of the books and the users. This cannot be automated, making it hard to upscale, and the analysis might be open to bias. Furthermore, using data captured on different time periods lead to different results.

Another way to discover similar titles is by adding standardised metadata. Most libraries use a classification for this purpose, which is standardised but rigid. Another option is using uncontrolled keywords that are flexible but lack standardisation. Wikipedia was used as ‘middle ground’: a standardised but very broad set of keywords. Adding Wikipedia pages to book records in the OAPEN Library is also reliant on external data, which must be provided by separate service. Furthermore, manual ‘culling’ of the results was necessary.

Both methods cannot be implemented completely automatically, rely on external services and need extra effort to scale up. This makes them less desirable for production. The solution described in this paper does not rely on external services but uses the strength of open access publishing: direct access to the contents of the documents.

Our goal is to find relevant open access titles, when a book or chapter has been selected. Relevant titles discuss the same concept or concepts that are closely connected. The algorithm is based on two assumptions: 1. The terms describing the themes of the title are frequently occurring in the text; 2. Books and chapters on the same subject use similar terms. In other words: if titles share relevant terms, they are connected. The number of shared relevant terms is an indication of the strength of the connection. Figure 2 displays the complete algorithm.

The next question is what terms to use. In this experiment, the terms are sets of three words – trigrams. In a text, the number of trigrams is relatively small – compared to bigrams – while they are more specific. This leads to a more ‘workable’ set of possible items. However, not all trigrams are relevant for our purpose, and therefore it is important to filter out the ones that are not needed.

The first set of trigrams to discard contains words that are too common: stop words. Examples are “a”; “able”; “about”; “above”, and almost 1,200 more words for the English language. Comparable sets of stop words for German and Dutch were also deployed.

The next set to filter out is trigrams that contain parts of words. When the contents of the books are converted to text, hyphens are converted to spaces, leading to trigrams such as “diff ere nt”, “inso fe rn” or “werkge legenhe id”. These are not three words, but just one.

Furthermore, trigrams that are specific to open access publishing or academic writing are discarded. These are descriptions of Creative Commons licenses, or terms that are quite common in academic books, but are meaningless in themselves, such as “pdf letzter zugriff”, “pdf zuletzt geprüft”, phd diss university” or “phd thesis university”.

Also, the part of references that only contain the publisher’s name are filtered out. For instance, the trigram “manchester university press” does not convey which title is cited. As Manchester University Press has published hundreds of titles on many different subjects, linking books using this term does not describe any subject related connection. Of course, this also applies many other academic publishers.

It is important to note that the terms to be excluded are a clearly visible part of the algorithm. This ensures maximum transparency: each person working with the algorithm has direct access to the ‘filtering terms’ and might choose to update them.

At the start, 12,224 titles in the OAPEN Library were selected. The selection was based on one criterium: language. The books and chapters were published in English, German, Dutch or a combination of these languages. Choosing these three languages was pragmatic: over 90 percent of the OAPEN Library collection is published in either English, German or Dutch, ensuring a sizable set of titles to analyse. Having a data set spanning multiple languages also enables possible connections between books in several languages. In one of the examples in section 4.2, we will find two closely connected books: an English language translation of a German book.

The first phase of data gathering was an attempt to download the full text of the titles. From each text, the most relevant trigrams were selected and lastly, for each title was determined if it could be matched to one of more other books or chapters. During this process, some texts could not be extracted, or no matching title could be found. This led to a dropout rate of around 10 percent, resulting in the 10,997 titles of the data set.

The data set is dominated by books (see Table 1); only 4% are chapters. Within the data set, English and German stand out, with a small percentage – 6% – of titles in Dutch or in multiple languages.

Each book or chapter in the data set is connected to one or more titles. The majority of the titles – over 7,000 – are closely related to 50 titles or less. Another 1,986 are connected to 100 titles or less. When the largest group is subdivided, it becomes clear that 4,498 books or chapters are closely connected to 20 titles of less. In other words, 40% of the titles.

Each title shares one or more trigrams with another publication. As is clearly visible in the histogram (Figure 3), most books and chapters are connected to 21 titles or more. Most of these connections vary in the number of shared trigrams. The number of shared trigrams is an indication of the strength of the connection: a higher number indicates a stronger connection.

These connections could be ranked. For instance, if a book is connected to 25 books – two books with three shared trigrams, five books with two shared trigrams and the rest with one shared trigram – these could be ranked first, second and third. However, we could also imagine several books that share a higher number of trigrams, where the first ranked titles share ten trigrams, the next six etc. Thus, the connections between the publications can be ranked, and their relative strength can be measured. This enables us to make specific selections, based on these parameters. The next section describes some examples.

This example is based on the book “Complexity, Security and Civil Society in East Asia” (Hayes & Yi, 2015), which discusses complex global problems such as urban insecurity, energy, and climate change. It shares three trigrams with four titles, one of them is the book “Loss and Damage from Climate Change” (Mechler et al., 2019). These four titles are part of the first rank. Moving on to the second rank, there are 29 tiles. One of those titles is “Louisiana’s response to extreme weather” (Laska, 2020). Furthermore, it shares one trigram with 217 titles, among them the book “Sustainable rice straw management” (Gummert et al., 2020); here the connection with insecurity and climate change seems weaker. However, the trigram both books share is “greenhouse gas emissions”.

Figure 4 displays the three ranks:

When looking at the connection between this book and the closely related books, the first question is what trigrams they share. These are listed in Table 2, Shared trigrams. The common theme connecting these books is quite clear: global warming and its effects.

However, the book “Complexity, Security and Civil Society in East Asia” does not only focus on climate change, and the trigrams reflect that. The most common trigrams are “civil society organizations” (occurring 78 times); “rok foreign policy” (occurring 57 times) and “world economic forum” (occurring 43 times). The first ‘shared’ trigram is “greenhouse gas emissions”, which is mentioned 29 times. The term “climate change adaption” is mentioned 20 times – the almost identical trigram “climate change mitigation” was counted 17 times. Lastly, “sea level rise” could be found 14 times.

It is also interesting to look at which trigrams could not be linked. Several of them are related to policy making, which became clear from the top three trigrams and several mentions of the Nautilus Institute for Security and Sustainability, a public policy think-tank. Furthermore, nuclear energy and energy security are also mentioned in several trigrams. The complete list can be found in section Appendix.

The previous section showed the titles related to one book. Another possibility is to examine groups of publications and their relations. What books are closely connected, and does their relative ‘distance’ display subtopics within a larger collection? Figure 5 shows a selection of books that share three trigrams or more. Each of the groups consists of closely connected books and chapters.

Randomly selecting titles based on the number of shared connections does not lead to very useful results. Starting with one book, it makes sense to search for related titles. In order to find more relevant results for groups, it is necessary to use additional metadata. In this case, the metadata of the OAPEN Library.

Using the metadata of the OAPEN Library enables us to search using several characteristics. In the next example, Figure 6 displays books published by Language Science Press. This publisher specialises in linguistics and all titles are part of a series; the colour of the cover denotes a series which helps to visualise the relations further. For instance, the green covers are part of the series “Studies in diversity linguistics”, and the dark blue covers indicate the “Computational models of language evolution” series. Moreover, the thickness of the connecting line is an indication of the number of shared trigrams.

Instead of focussing on a single publisher, we could also look at the open access titles that received financial support from the same funder. If the funder has an underlying policy regarding the titles – see for instance Rieck (2019) – is that reflected in the publications? Figure 7 displays books funded by the Austrian Science Fund (FWF). Here, several smaller groups of closely connected books are noticeable.

Furthermore, the two titles in the bottom right are translations: “Revolution and transition : Cultural policy in Bulgaria, 1989–2012” (Alexandrov, 2017a) and “Wende und Übergang : Die Kulturpolitik Bulgariens, 1989–2012” (Alexandrov, 2017b). The algorithm is capable of connecting books across languages. More on translations in the next section.

The graphics in this section were created using NodeXL (Smith et al., 2010). The data set and the algorithm in the R language is available at https://doi.org/10.17026/dans-xbm-qr5e.

The connection between the two translated books in the set of FWF funded titles is not a coincidence. Within the data set, at least 15 “translated couples” could be found. This might seem counterintuitive: the algorithm is based on finding exact trigrams, and one would expect translations to use different words to describe the same concepts. However, the analysis of several sets of translated books that share nine or more trigrams shows they often share English language terms, such as “adaptive cruise control” (Maurer et al., 2015, 2016); “labour force survey” (Holtslag et al., 2012, 2013) or “deep packet inspection” (Sprenger, 2015a, 2015b). Nevertheless, the shared terms do not have to be restricted to English, such as “graf leo thun” (Aichner & Mazohl, 2017a, 2017b). Additionally, web addresses also function as a language agnostic identifier. See for instance “http://www.siebenbuerger.de zeitung” (Hermanik, 2016a, 2016b) or “http://www.minfin.bg bg” (Alexandrov, 2017a, 2017b).

The previous sections described some of the possible applications of the trigram algorithm, based on a single books or groups of titles. What are possible use cases for the stakeholders involved? The first use case is based on the connections surrounding a single title. As discussed in the introduction, this can be used to create a recommender system. For each title, the recommender system might display titles ranked first to third. The selection could also be refined by the number of titles: in the example of section 4.1, the number of third ranked titles linked to the book is 217, which is possibly too much for a single recommendation.

Creating benchmarks for publishers would be another use case. Here, the goal is comparing usage data of a set of comparable titles to a publication. By selecting all connected titles and collecting usage data it is possible to establish the average usage for this particular publication. This can be used as benchmark. Again, the number of titles to include can be varied by selecting only higher ranked titles.

Libraries might be interested in creating a collection of connected titles. Using the metadata such as keywords or classification creates a core set of titles, which can be expanded by selecting connected titles. Once more, the differences in ranking help to determine the extensiveness of the collection. A similar approach could also be used by researchers, looking for related titles to be used for citation or usage analysis.