2026-10-22: The 30th International Conference on Theory and Practice of Digital Libraries (TPDL) 2026 Trip Report

This year, the 30th International Conference on Theory and Practice of Digital Libraries (#TPDL2026) was held in the beautiful town of Faro, Portugal from September 22-25, 2026. We, Himarsha Jayanetti and Skanda Siva, attended the conference in person to present our paper titled “Nobody Truly Agrees on Sentiment: Humans, Bespoke Tools, and LLMs Struggle with Social Media Texts”, co-authored with our advisors, Michele C. Weigle, Michael L. Nelson, and Shuai Hao, all from the Department of Computer Science at Old Dominion University.

InfoVis-DL Workshop

The first day of the conference was dedicated to workshops and the Doctoral Consortium. We attended the InfoVis-DL 2026 workshop titled “Information Visualization in Digital Libraries: Multi-modal Applications to Cultural Heritage Collections”, which explored how information visualization can help libraries, archives, museums, and other cultural heritage institutions make complex digital collections more accessible and meaningful to users.

📍 #TPDL2026 in Faro, Portugal 🇵🇹 Happening now: InfoVis-DL — Information Visualization in Digital Libraries: Multi-modal Applications to Cultural Heritage Collections. Excited for a great workshop and discussions! @webscidl.bsky.social @phonedudemln.bsky.social @weiglemc.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 10:02 AM · Sep 22, 2026

The workshop brought together several interesting perspectives on visualizing cultural heritage collections. “From Inscription to Information: A Multi-modal Visualization Pipeline for Endangered Graffiti Heritage at Apostolos Andreas Monastery, Cyprus” presented a fascinating effort to digitally document endangered historical graffiti using archival photographs, photogrammetry, image enhancement, 3D modeling, and other multi-scale techniques. Their work demonstrated how visualization can become more than a way of presenting a collection and how it can serve as a research and preservation tool when the original material is deteriorating or has already disappeared. One statement from the presentation that I particularly remember was “Visualization was not the last stage of this research; it is what provided the means for this research”. The second presentation, “Tracing Early Digital Art: Coordinated Visual Exploration of Digital Art Archives” approached visualization from a different direction, bringing together archival records about artists, exhibitions, institutions, and communities and allowing users to explore their relationships through interconnected network, temporal, and geographic views. An especially interesting point from the discussion was that visualizations are not neutral: what appears to be an absence or an isolated artist in a visualization may simply reflect gaps in the underlying archival record. The third presentation introduced DAViD Lab, Portugal’s first laboratory dedicated to the digitisation and advanced visualisation of historical archives. The lab plans to combine AI-assisted exploration with large-scale displays, AR and VR, spatial sound, touch and gesture-based interaction, and other immersive technologies, with potential applications ranging from cultural heritage and historical propaganda collections to transportation archives. The presentations led to a lively discussion about accessibility, open research, interoperability, storytelling, public engagement, and how visualization can help bring digital collections closer to both researchers and the general public. Together, the presentations sparked an interesting discussion about how visualization can go beyond simply displaying digitized material to help people explore, interpret, and connect with cultural heritage collections. 

Main Conference: Day 1

The first day of the main conference began with the keynote by María Poveda-Villalón, Associate Professor at Universidad Politécnica de Madrid. Her talk was titled “Knowledge Beyond Documents: FAIR Ontologies for Humans and Machines” and she explained how to make ontologies and vocabularies FAIR (Findable, Accessible, Interoperable, Reusable) across the full development lifecycle. She emphasized the importance of making metadata and ontologies machine-actionable for agents and LLMs. One key takeaway was that FAIRness should be considered throughout the entire development process, rather than only after publication. She also highlighted the importance of standard vocabularies, persistent URIs, and interoperable formats such as RDF/OWL, along with lightweight, iterative development practices and automated FAIR checks. During her talk, she also introduced FOOPS! (FAIR Ontology Pitfall Scanner), a tool that automates FAIR assessments and helps developers identify areas where their ontologies can be improved.

#TPDL2026 Main Conference - Day 1 kicks off with the first keynote! 🎤 “Knowledge Beyond Documents: FAIR Ontologies for Humans and Machines” by María Poveda-Villalón @webscidl.bsky.social, @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 5:02 AM · Sep 23, 2026

The first session, “Data Management Plans and FAIR Assessment”, included three presentations exploring different aspects of machine-actionable (a term I learned here, distinct from machine-readable) data management and FAIR evaluation. The first paper, “Community Metrics and Automated Assessment of Machine-Actionable Data Management Plans”  by Tomasz Miksa presented work from the OSTrails project. The presentation highlighted a common problem with traditional Data Management Plans (DMPs) where researchers prepare them to satisfy funder requirements but receive little useful feedback. He introduced a catalog of 94 community-driven metrics for automating DMP evaluations and providing researchers with feedback. The discussion raised questions about whether automated checks could verify that researchers actually followed through on the commitments made in their initial DMPs. Anna Lehmann-Gnirck presented the  second paper of the session, “It’s FAIR and generic?! A Systematic Analysis for FAIR Assessment Tools”, which examined how different automated tools assess the FAIRness of research data. An interesting observation was that the same dataset can receive different FAIR scores depending on the tool used, because each tool selects and implements assessment criteria differently. Anna emphasized that automated assessments are generally better at checking technical aspects, such as persistent identifiers and metadata availability, than determining whether data is genuinely understandable and reusable within a particular research community. She proposed combining generic technical assessments with domain-specific semantic evaluations. This led to an important question: rather than simply asking whether a dataset scored 70% or 90% on a FAIR assessment, shouldn't we first ask what that score actually measures? The final paper of the session, “Enhancing Machine-Actionable DMPs with Policy Entities Expressed in ODRL” by Diamantis Tziotzios, returned to the OSTrails project and addressed another limitation of machine-actionable DMPs. Although DMPs can now represent much of their information in structured formats, policies governing data access and usage are often still written as free text. The authors explored using ODRL (Open Digital Rights Language), a W3C standard, to formally represent permissions, prohibitions, and obligations. This session offered very interesting perspectives on machine-actionable research data management: defining what can be automatically evaluated, understanding what FAIR assessment scores actually tell us, and representing policies in ways that machines can interpret and act upon.

#TPDL2026 Session 1: Data Management Plans and FAIR Assessment is happening now! Chaired by Ralf Krestel. The papers are available to download: tpdl2026.ualg.pt/program/

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 6:41 AM · Sep 23, 2026

A social event was held that evening, featuring a guided tour of the Museu de Faro followed by a walking tour of Faro Old Town.

A few captures from the #TPDL2026 social event yesterday evening 📸🇵🇹 A tour of the Faro Museum, followed by a walk through the beautiful Old Town of Faro. A lovely way to wrap up the day, made extra special with our little one by our side 💕 @webscidl.bsky.social, @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 4:47 AM · Sep 24, 2026

Main Conference: Day 2

The second keynote of the conference was by Gianmaria Silvello, Professor at the Department of Information Engineering, University of Padua. His talk was titled, “Transitive Credit Allocation for Heterogeneous and Dynamic Scholarly Graphs”. Traditional citation metrics give credit to directly cited papers, but often overlook the foundational research, datasets, and software that made those papers possible. He introduced a transitive citation credit model in which a portion of the credit received by a paper is passed backward through its citation network, allowing earlier contributions to receive recognition. The model was evaluated on a large scholarly citation graph containing approximately 38 million papers, demonstrating how accounting for indirect contributions can significantly change researcher rankings. One particularly interesting example was a 1955 paper on antibody sequences that had received only 19 direct citations, but accumulated more than 5,000 units of credit when its indirect influence was considered. The keynote also raised a timely question about credit attribution in the age of LLMs: when AI-generated answers rely on knowledge derived from numerous publications and datasets, how can we recognize the original contributors? Reading about this work reminded Dr. Nelson of a quote he said he never misses a chance to mention: “Everyone Who Bought One of Those 30,000 Copies Started a Band”. The Velvet Underground, one of the most influential rock bands of the 1960s and early 1970s, although they didn’t sell many records, their music inspired many of those who listened to it to start their own bands. The quote nicely captures the message of the keynote: the true influence of a work is not always reflected in its immediate recognition, but in the many contributions it inspires over time.

#TPDL2026 Main Conference - Day 2 kicks off with the second keynote 🎤 “Transitive Credit Allocation for Heterogeneous and Dynamic Scholarly Graphs” By Gianmaria Silvello, Professor at the Department of Information Engineering, University of Padua @webscidl.bsky.social, @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 5:00 AM · Sep 24, 2026

We attended the “LLMs and Multimodal Models for Knowledge Discovery” session, which had four presenters on how AI models can support information retrieval and knowledge discovery across historical image collections, scientific literature, and visual data. First presentation was “Using Hierarchical Controlled Vocabularies to Understand CLIP Retrieval Failures in Historical Photo Collections”, by Ratan Sebastian on why CLIP sometimes struggles to retrieve relevant images from historical photograph archives. The authors investigated whether retrieval failures occur because a model fails to associate a query with relevant images or because the images themselves are not represented as a coherent group. One interesting finding was that poor query-image alignment was a particularly important source of retrieval failure, while fine-tuning could improve performance for broader, more abstract concepts. The work highlighted the value of understanding why a retrieval model fails rather than relying only on overall performance scores. The second paper, “MARIO: Multi-Agent Retrieval and Iterative Reasoning for Knowledge-Augmented Generation in Scientific Question Answering” by Maodi Hu introduced a multi-agent approach to answering complex scientific questions. Instead of relying on a single retrieval step, MARIO combines knowledge retrieval with iterative reasoning, allowing different agents to gather and evaluate evidence. An important aspect of the approach was connecting generated answers to supporting scientific sources, making the reasoning process more traceable.

The third paper, “GeoSearch: Transforming Geoscience Literature Discovery for Chinese Geoscientists with AI” by Feng Zhou presented an AI-based approach to helping researchers discover relevant geoscience literature, addressing the challenges of finding useful information within a large and specialized body of scientific publications. He highlighted how fragmented sources, inconsistent metadata, and varying terminology make it challenging to find and connect relevant scientific knowledge. The final presentation of the session, “Do you ‘see’ what I ‘see’? A Multi-panel Visual Question and Answer Dataset for Large Language Model Chart Analysis”, by Daniel Evans raised another interesting question: Can an AI model truly understand a scientific figure when the answer requires connecting information across multiple charts? Reading individual labels or identifying trends in one chart is different from interpreting relationships between several panels. All four presentations approached AI from different perspectives but shared a common concern: retrieving or generating an answer is only part of the problem; understanding why it is correct, and where these models fail, is equally important. I think this is a valuable consideration for almost any research project involving AI today.

#TPDL2026 Session 6 · LLMs and Multimodal Models for Knowledge Discovery is happening now! Chaired by Anett Hope. The papers are available to download: tpdl2026.ualg.pt/program/

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 6:37 AM · Sep 24, 2026

The conference dinner took place that evening at Elemento on Faro Island. We reached there by ferry across the beautiful Ria Formosa lagoon. The scenic ferry ride was a highlight of the evening, followed by an enjoyable dinner featuring plenty of Portuguese sangria, fresh seafood, and an opportunity to socialize with fellow attendees. During the dinner, the organizers also announced the Best Paper Award, with “ReproScore: Separating Readiness from Outcome in Research Software Reproducibility Assessment” by Sheeba Samuel et al. being recognized as the winning paper.

At the #TPDL2026 conference dinner! ⛴️🌊 A beautiful ferry ride across the Ria Formosa to Elementos Faro for dinner by the waterfront ✨ A great way to wrap up another conference day! @webscidl.bsky.social @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 6:19 PM · Sep 24, 2026

Main Conference:  Day 3

Adam Jatowt, a professor at University of Innsbruck, delivered the third and final keynote of the conference, titled “Temporal Intelligence in Language Technologies: Modeling, Retrieval, and Reasoning”. He discussed the importance of integrating time (timestamps, dates, relative expressions) into language-model representations and training to improve factual correctness and document-centered question answering. He emphasized that a fact can be correct at one point in time but incorrect at another. For example, answering a question about who held a particular position requires understanding not just the information itself, but also the time period to which it applies. This becomes particularly challenging when documents contain outdated information or refer to events using relative expressions such as “last year” or “two months ago.” The keynote highlighted why language technologies need to go beyond recognizing temporal expressions and develop the ability to reason about how information and facts change over time.

The final day of #TPDL2026 is today! 🎤 Day 3 begins with the conference’s third keynote: “Temporal Intelligence in Language Technologies: Modeling, Retrieval, and Reasoning” by Adam Jatowt @webscidl.bsky.social @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 4:44 AM · Sep 25, 2026

The final session of TPDL 2026, “AI, Human Judgement, and Public Information Practices”, featured work exploring how people use, evaluate, and interact with AI and information systems.

Luana Ponte presented “A Complementary Tool for a Sensitive Domain: How Users Seek Health Information Through LLMs”, examining how people incorporate LLMs into their health information searches. The study found that users generally treat LLMs as a complement to traditional search engines rather than a replacement, particularly when seeking clarification on health-related questions. The discussion also raised an important question about how AI interfaces could better encourage users to verify information in high-stakes domains. In the session's only full paper, “Speaker Mining – FAIR Data on Public Broadcasts for Question Answering”, Tim Wittenborg presented efforts to integrate fragmented German public broadcasting metadata into a structured, reusable knowledge base. By linking television episodes and guest appearances to Wikidata, the researchers demonstrated how structured broadcast data could support questions about media representation, including who appears on talk shows and how frequently. The presentation also highlighted the substantial manual effort required to reconcile records across different sources. Maja Kuhar and Adam Roegiest presented “Adoption, Attitudes, and Apprehensions: A Survey of Librarians on GenAI”, focusing on librarians in Slovenia. Their findings showed that librarians recognize GenAI's potential for tasks such as cataloging, reference services, and administrative work, but remain cautious about accuracy, privacy, and overreliance on these tools. The presenters emphasized that GenAI should support librarians rather than replace the human expertise and interpersonal interactions central to their work.

The final session of #TPDL2026! Session 8 - AI, Human Judgement, and Public Information Practices The papers are available to download: tpdl2026.ualg.pt/program/ @webscidl.bsky.social @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 5:48 PM · Sep 25, 2026

The final session concluded with our paper, “Nobody Truly Agrees on Sentiment: Humans, Bespoke Tools, and LLMs Struggle with Social Media Texts”. In this work, we compared sentiment judgments from human annotators, traditional sentiment analysis tools, and LLMs to examine how consistently they interpret social media text. Our results showed that even experienced human annotators often disagreed on sentiment, particularly when distinguishing neutral expressions from positive or negative ones. We also found that simplifying the task to negative versus non-negative or positive versus non-positive generally improved agreement. Among the automated systems, Twitter-RoBERTa aligned most consistently with the human consensus, achieving a Cohen's kappa of 0.73 for three-class classification and 0.88 for negative versus non-negative classification. Although the LLMs often agreed strongly with one another, their agreement did not necessarily reflect human judgments. Our findings highlighted two important considerations: sentiment is inherently subjective, making human disagreement an important part of evaluation, and domain-specific models can still outperform general-purpose LLMs on tasks closely aligned with their training data. The audience discussion also raised a valuable question about whether we should treat human disagreement as useful information rather than reducing it to a single majority-vote ground truth.

The conference concluded with closing remarks by José Barateiro, who thanked the organizers, speakers, volunteers, and attendees for making TPDL 2026 a success. Gianmaria Silvello then announced that TPDL 2027 will take place on San Servolo Island in Venice, Italy! We had such a wonderful time at TPDL 2026 and hope to see everyone again next year in beautiful Venice for more inspiring research, great conversations, and familiar faces!

And that’s a wrap on #TPDL2026! Closing remarks by José Barateiro, followed by Gianmaria Silvello announcing #TPDL2027 in Venice, Italy! 🇮🇹 - inviting everyone to return next year and, of course, to start working on those papers! @webscidl.bsky.social, @skandasiva.bsky.social

[image or embed]

— himarshaj.bsky.social (@himarshaj.bsky.social) 6:29 PM · Sep 25, 2026

For a broader perspective on the conference and its history, take a look at “Thirty Years of Digital Libraries: What ECDL/TPDL Tells Us About the Field, and Where We Go from Here” by Gianmaria Silvello. The paper looks back at 30 years of ECDL/TPDL and examines how the field has evolved. Old Dominion University is well represented in this analysis, a connection that continues through our participation in TPDL 2026. 

For another glimpse into the WS-DL group’s TPDL experiences, see our TPDL 2022 trip report, written after Emily and Himarsha attended the conference that year.

Acknowledgements

We would like to thank Drs. Michele Weigle and Michael Nelson for supporting Himarsha’s travel to TPDL 2026. We also thank Dr. Shuai Hao, the Computer Science Graduate Society (CSGS), and SEES travel awards for partially supporting Skanda’s travel. We are grateful for their support, which made it possible for us to share such a memorable conference experience together in beautiful Faro, Portugal.

- - Himarsha R. Jayanetti and Skanda Siva