Note: This is a guest post by Michael LeBlanc, Senior Manager – Library Applications at UBC Information Technology.

I was about five years and three hundred pages into my doctoral dissertation when I stumbled across Marcus Beck’s famous analysis of dissertation length at the University of Minnesota. Beck, then a doctoral student, published a 2013 blog post on average dissertation lengths at the university, followed by a revision in 2014 including Master’s theses. He seemed to be a kindred spirit: a pen-weary graduate student also wondering if he had written far too much. I was in luck! His research suggested I had not strayed far from the landing zone, with a history dissertation median of just under 300 pages.

Of course, I wasn’t writing a history dissertation (technically, it was in education), and I wondered whether UBC’s numbers mirrored Minnesota’s. However, UBC’s repository didn’t list page or word-count information for dissertations and theses, so I tabled the idea. Two years later, I finished and submitted my draft—and my committee was incredible. They were supportive, thoughtful, and possibly a bit tired. The dissertation was 466 pages long.

Accessing UBC’s Graduate Research

When I finished my doctorate, I returned to the question of what was a long dissertation. I wondered whether the metric had changed over time—after all, the dissertation has a history too. Ruano-Borbalan (2022) writes that the dissertation only emerged as an original, written piece of doctoral research in the eighteenth century. Beck’s analysis focused on theses and dissertations since 2007.

UBC graduate students deposit their theses and dissertations to cIRcle, the university’s open access digital repository for material created by the UBC community and its partners. Open Collections is a discovery interface for cIRcle and other UBC repositories, with holdings ranging from historical BC newspapers and photographs to the theses and dissertations that interest me. Its metadata makes it possible to find relevant items by title, author, supervisor (sometimes), description, degree, program, and more. However, returning to my original question, it doesn’t tell us about pages or words per document.

Building the Workbench

So I decided to write a tool called the Dissertation Workbench (co-developed with GenAI), available on GitHub under an MIT License. It communicates with Open Collections’ public APIs to query document metadata and full text. It began as a simple application, limited to EdD dissertations in Educational Studies, counting their pages and words.

However, I realized I could generate additional document information as the Workbench gathered these basic statistics. What began as a question about dissertation length became an exploration of graduate research at UBC. Who were the supervisors? Which methods did students use? How had topics changed over time? Others have explored library metadata as a primary information source, such as Manolis Peponakis (2013) and Gooding, Terras, and Ames (2025). I decided to extend the Workbench to identify supervisors, methodologies, concepts, citations and topics.

These new attributes would also have practical applications for discovery. Thinking back to my experience as a graduate student, I turned to Open Collections more than once to locate research by topic and methodology. I was still bothered by a dissertation on educational technology in the Yukon—located late one night and never surfacing again (until it reappeared in the Workbench). To make these connections, I needed to extract various facets of information from Open Collections. I used a variety of strategies, from pattern matching (regular expressions) to local artificial intelligence; I’ll briefly discuss them below.

Finding Supervisors

Open Collections does not always contain supervisor metadata for older theses and dissertations, particularly those submitted before 2018. The Workbench relies on sentimentality to fill the gaps, searching Acknowledgements sections. This is an imperfect mechanism because not every document includes acknowledgments; a brief survey suggests this practice varied across disciplines and time. Concerned by the ethical problems of inaccurate attribution, I chose pattern matching rather than LLM inference, looking for instances of “my supervisor,” “my advisor…”, etc.

Methodology Detection

Methodology detection also employs pattern matching, applying a curated dictionary of methods to document abstract and title metadata. This lightweight approach relies on methodology language rather than methodological indicators in the full text. If an author used interviews but did not mention them in the abstract or title, that method would be missed.

Concepts and Topics

I turned to PatternRank for keyword identification, labelled as “concepts” in the Workbench (Schopf et al., 2022). It uses grammatical patterns to extract phrases from the abstract and then applies a neural language model (SBERT) to determine how closely they map to the document’s overall meaning. For connections between documents, the Workbench uses BERTopic and an embedding language model (AllenAI/Specter2) to create topical groupings. For example, my dissertation belongs to the topic “Educational Culture, Internationalism, and Social Reform” with six others.

By mapping dissertations to topics and topics to time, we can see how research interests have evolved. I want to be careful, though, because topic naming involves some subjectivity. BERTopic produces a bag of words associated with each topic, and someone must distill those disjointed words into a label. If poorly done, a topic can seem like something it is not.

Locating References

Lastly, the Workbench locates references in the corpus of theses and dissertations processed so far, showing which works lean toward canonical status. Extracting references in different formats, from different eras, is a complex problem. Meuschke et al. (2023) identify Grobid as a capable tool for this task, which I implemented as a utility in the Workbench. Grobid parses PDFs and provides the resulting citations as a structured data document. The Workbench then links those citations to the thesis or dissertation. The pipeline also includes fallback parsers if Grobid’s citation count is unexpectedly low.

Discovering Theses and Dissertations in the Workbench

The Workbench offers enriched metadata and collection-level insights through four interfaces. The Document Explorer lets researchers view theses and dissertations with Workbench-derived metadata: page and word counts, concepts, topics, supervisors, and citations. The Citation Explorer shows foundational works—which sources do UBC’s student researchers reference again and again? The Person Explorer maps supervisors and committee members to student research. And finally, the Analytics Dashboard charts page and word counts, methodology use, and topic trends over time.

The Workbench remains in beta. It began with 420 dissertations, and it has now parsed about 14,000 documents. I have since re-engineered it to support more—hopefully the full collection of 56,000 theses and dissertations in cIRcle. As the Workbench’s scope grows, its metadata extraction and document classification processes will need refinement. If you have thoughts, suggestions, or corrections, please contact me through cIRcle.

The Workbench Offers Clues, Not Conclusions

The Workbench illustrates how cIRcle and Open Collections contribute to knowledge generation at a meta-level, through open access to thesis and dissertation metadata. By using that metadata and drawing on source documents, the Workbench creates new information about student research at UBC. The Workbench wouldn’t work without that foundation.

I hope that researchers, and students in particular, can use the Workbench to uncover avenues of inquiry or resources to support their work. Alternatively, a look at simple page counts may inspire them to refine the parameters of their project. Supervisors and graduate program advisors may look to the Analytics Dashboard for a longer view of how methods and interests in their program have shifted. However, I would caution against treating Analytics Dashboard metrics as the truth. The Workbench remains a work in progress, but it is also open source. Examine the metadata and reports, and the code that produces them. Correct it, if necessary, and take it in directions that are personally meaningful.

In my case, about halfway through developing the Workbench, I decided to parse Master’s theses too—wondering about my own from 22 years ago. The Workbench quickly showed me that I held first place again. This time it was for brevity in History at 34 pages. So I dusted off the bound copy on my bookshelf, and sure enough, the Workbench was right.

Perhaps there is something to be said about things averaging out over time.

References

Gooding, P., Terras, M., & Ames, S. (2025). Library catalogues as data: Research, practice and usage (1st ed.). Facet Publishing.

Meuschke, N., Jagdale, A., Spinde, T., Mitrović, J., & Gipp, B. (2023). A benchmark of PDF information extraction tools using a Multi-task and Multi-domain evaluation framework for Academic documents. In I. Sserwanga, A. Goulding, H. Moulaison-Sandy, J. T. Du, A. L. Soares, V. Hessami & R. D. Frank (Eds.), Information for a better world: Normality, virtuality, physicality, inclusivity (pp. 383-405). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-28032-0_31

Peponakis, M. (2013). Libraries’ metadata as data in the era of the semantic web: Modeling a repository of master theses and PhD dissertations for the web of data. Journal of Library Metadata, 13(4), 330-348. https://doi.org/10.1080/19386389.2013.846618

Ruano‐Borbalan, J. (2022). Doctoral education from its medieval foundations to today’s globalisation and standardisation. European Journal of Education, 57(3), 367-380. https://doi.org/10.1111/ejed.12522

Schopf, T., Klimek, S., & Matthes, F. (2022). PatternRank: Leveraging pretrained language models and part of speech for unsupervised keyphrase extraction. Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management, 243–248, https://doi.org/10.5220/0011546600003335.