kb

Finding Near Duplicates in Digital Heritage Collections

~ Victor ~ Leave a comment

[This post is based on Senna Lindner’s Information Science Master thesis]

Digital libraries contain millions of documents, scans, newspapers, and books. As these collections continue to grow, so does a less visible problem: near duplicates. These are files that are not exact copies, but are almost identical, for example because the same document has been scanned twice, or because a thesis has been downloaded again with only a different cover page. Detecting such duplicates is difficult, but important for maintaining large digital collections.

Example of near duplicate scans

In his Master’s thesis, Senna Lindner investigated how near duplicate documents can be detected in the collections of the Koninklijke Bibliotheek (KB), the Dutch National Library. The research compared four existing duplicate detection techniques: the text-based methods SimHash and MinHash, and the image-based methods pHash and dHash. Rather than evaluating these methods individually, the study also explored combinations of methods.

The results showed that there is no single best method for all types of documents. For text-rich PDF theses, MinHash alone achieved perfect detection performance. For scanned documents, however, the best results came from combining text and image-based approaches, achieving an F1 score of over 93%. These hybrid approaches were more accurate, although they also required more computing time.

To understand how such a system could be used in practice, interviews were conducted with staff members of the KB. While interviewees acknowledged that near duplicates exist, they emphasized that usability and workflow integration are as important as detection performance. Participants preferred a human-in-the-loop approach, where the system suggests possible duplicates and staff members verify the results before any action is taken. High recall, finding as many duplicates as possible, was generally considered more important than perfect precision.

The thesis demonstrates that combining multiple detection methods can significantly improve the identification of near duplicates in digital heritage collections. At the same time, successful adoption depends not only on algorithms, but also on designing tools that are transparent, efficient, and fit existing archival practices.

The Github link to the code and data used in Senna’s thesis: https://github.com/sennalindner/Near-duplicate-detection

Senna Lindner. Near Duplicate Detection in Digital Libraries. Master’s Thesis, Vrije Universiteit Amsterdam, in collaboration with the Koninklijke Bibliotheek, 2026.

Master_IS_Research_Thesis_Senna_Lindner-7Download

Share This:

Automating Authorship Attribution

~ Victor ~ Leave a comment

[This blog post was written by Nizar Hirzalla and describes his VU Master AI project conducted at the Koninklijke Bibliotheek (KB), co-supervised by Sara Veldhoen]

Authorship attribution is the process of correctly attributing a publication to its corresponding author, which is often done manually in real-life settings. This task becomes inefficient when there are many options to choose from due to authors having the same name. Authors can be defined by characteristics found in their associated publications, which could mean that machine learning can potentially automate this process. However, authorship attribution tasks introduce a typical class imbalance problem due to a vast number of possible labels in a supervised machine learning setting. To complicate this issue even more, we also use problematic data as input data as this mimics the type of available data for many institutions; data that is heterogeneous and sparse of nature.

Inside the KB (photo S. ter Burg)

The thesis searches for answers regarding how to automate authorship attribution with its known problems and this type of input data, and whether automation is possible in the first place. The thesis considers children’s literature and publications that can have between 5 and 20 potential authors (due to having the same exact name). We implement different types of machine learning methodologies for this method. In addition, we consider all available types of data (as provided by the National Library of the Netherlands), as well as the integration of contextual information.

Furthermore, we consider different types of computational representations for textual input (such as the title of the publication), in order to find the most effective representation for sparse text that can function as input for a machine learning model. These different types of experiments are preceded by a pipeline that consists out of pre-processing data, feature engineering and selection, converting data to other vector space representations and integrating linked data. This pipeline shows to actively improve performance when used with the heterogeneous data inputs.

Implemented neural network architectures for TFIDF (left) and Word2Vec (right) based text classification

Ultimately the thesis shows that automation can be achieved in up to 90% of the cases, and in a general sense can significantly reduce costs and time consumption for authorship attribution in a real-world setting and thus facilitate more efficient work procedures. While doing so, the thesis also finds the following key notions:

  1. Between comparison of machine learning methodologies, two methodologies are considered: author classification and similarity learning. Author classification grants the best raw performance (F1. 0.92), but similarity learning provides the most robust predictions and increased explainability (F1. 0.88). For a real life setting with end users the latter is recommended as it presents a more suitable option for integration of machine learning with cataloguers, with only a small hit to performance.
  2. The addition of contextual information actively increases performance, but performance depends on the type of information inclusion. Publication metadata and biographical author information are considered for this purpose. Publication metadata shows to have the best performance (predominantly the publisher and year of publication), while biographical author information in contrast negatively affects performance.
  3. We consider BERT, word embeddings (Word2Vec and fastText) and TFIDF for representations of textual input. BERT ultimately grants the best performance; up to 200% performance increase when compared to word embeddings. BERT is a sophisticated language model with an applied transformer, which leads to more intricate semantic meaning representation of text that can be used to identify associated authors. 
  4. Based on surveys and interviews, we also find that end users mostly attribute importance to author related information when engaging in manual authorship attribution. Looking more in depth into the machine learning models, we can see that these primarily use publication metadata features to base predictions upon. We find that such differences in perception of information should ultimately not lead to negative experiences, as multiple options exist for harmonizing both parties’ usage of information.
Summary of the final performances of the best performing models from the differing implemented methodologies

Share This:

Who uses DBPedia anyway?

~ Victor ~ Leave a comment

[this post is based on Frank Walraven‘s Master thesis]

Who uses DBPedia anyway? This was the question that started a research project for Frank Walraven. This question came up during one of the meetings of the Dutch DBPedia chapter, of which VUA is a member. If usage and users are better understood, this can lead to better servicing of those users, by for example prioritizing the enrichment or improvement of specific sections of DBPedia Characterizing use(r)s of a Linked Open Data set is an inherently challenging task as in an open Web world, it is difficult to know who are accessing your digital resources. For his Msc project research, which he conducted at the Dutch National Library supervised by Enno Meijers , Frank used a hybrid approach using both a data-driven method based on user log analysis and a short survey of know users of the dataset. As a scope Frank selected just the Dutch DBPedia dataset.

For the data-driven part of the method, Frank used a complete user log of HTTP requests on the Dutch DBPedia. This log file (see link below) consisted of over 4.5 Million entries and logged both URI lookups and SPARQL endpoint requests. For this research only a subset of the URI lookups were concerned.

As a first analysis step, the requests’ origins IPs were categorized. Five classes can be identified (A-E), with the vast majority of IP addresses being in class “A”: Very large networks and bots. Most of the IP addresses in these lists could be traced back to search engine

indexing bots such as those from Yahoo or Google. In classes B-F, Frank manually traced the top 30 most encounterd IP-addresses, concluding that even there 60% of the requests came from bots, 10% definitely not from bots, with 30% remaining unclear.

The second analysis step in the data-driven method consisted of identifying what types of pages were most requested. To cluster the thousands of DBPedia URI request, Frank retriev

ed the ‘categories’ of the pages. These categories are extracted from Wikipedia category links. An example is the “Android_TV” resource, which has two categories: “Google” and “Android_(operating_system)”. Following skos:broader links, a ‘level 2 category’ could also be found to aggregate to an even higher level of abstraction. As not all resources have such categories, this does not give a complete image, but it does provide some ideas on the most popular categories of items requested. After normalizing for categories with large amounts of incoming links, for example the category “non-endangered animal”, the most popular categories where 1. Domestic & International movies, 2. Music, 3. Sports, 4. Dutch & International municipality information and 5. Books.

Frank also set up a user survey to corroborate this evidence. The survey contained questions about the how and why of the respondents Dutch DBPedia use, including the categories they were most interested in. The survey was distributed using the Dutch DBPedia websitea and via twitter however only attracted 5 respondents. This illustrates

the difficulty of the problem that users of the DBPedia resource are not necessarily easily reachable through communication channels. The five respondents were all quite closely related to the chapter but the results were interesting nonetheless. Most of the users used the DBPedia SPARQL endpoint. The full results of the survey can be found through Frank’s thesis, but in terms of corroboration the survey revealed that four out of the five categories found in the data-driven method were also identified in the top five resulting from the survey. The fifth one identified in the survey was ‘geography’, which could be matched to the fifth from the data-driven method.Frank’s research shows that although it remains a challenging problem, using a combination of data-driven and user-driven methods, it is indeed possible to get an indication into the most-used categories on DBPedia. Within the Dutch DBPedia Chapter, we are currently considering follow-up research questions based on Frank’s research.

Share This: