Mat Kelly

ws-dl.blogspot.com2026-10-19: FidEx and Wayback QA, Two Tools That Assist With Identifying Why A Resource Failed To Load During ReplayThe Web Science and Digital Libraries Research Group at Old Dominion University.

37m

Just gave an MLIS lecture on LAM history (as a comp/info scientist) and still feel that I've a lot to learn.

Mat Kelly reposted

Travis Reid

@treid803.bsky.social

10h

In this blog post, I summarized what I have learned about the FidEx and Wayback QA tools and how these tools can help with identifying why a sub-resource failed to load during replay. @webscidl.bsky.social ws-dl.blogspot.com/2026/10/2026...

Mat Kelly reposted

Lucy Li

@lucy3.bsky.social

The information school here is also hiring! "We especially encourage applications from ... the intersection of information, people, technology, and society..." jobs.wisc.edu/jobs/assista...

jobs.wisc.eduAssistant / Associate / Full Professor - Madison, Wisconsin, United StatesCurrent Employees: If you are currently employed at any of the Universities of Wisconsin, log in to Workday to apply through the internal application process.Job Category:FacultyEmployment Type:Regula...

Lucy Li@lucy3.bsky.social

We are hiring!! "Artificial intelligence, broadly construed. This includes embodied AI and robotics, machine learning and the systems that support it, and AI developed for problems in science, engineering, medicine, education, and society." jobs.wisc.edu/jobs/assista...

jobs.wisc.eduAssistant / Associate / Full Professor of Computer Sciences - Madison, Wisconsin, United StatesCurrent Employees: If you are currently employed at any of the Universities of Wisconsin, log in to Workday to apply through the internal application process.Job Category:FacultyEmployment Type:Regula...

At the "Text and Language Processing" session at #jcdl2026, Shirin Shujaa from RMIT Univ. describes their controlled study of hallucination behavior under context compression. They introduce Attention-Based Semantic Preservation, a method that retains tokens based on structural integrity.

Their results imply that it's cheaper to peform the process using a prompt. It's both less expensive and more accurate. VusURLX (w/ GPT 5-mini) had the highest F1 score for extracting URLs and retaining context. #jcdl2026

They built a benchmark of text-layer scholarly URLs w/ their context. Their work was supported by IMLS ♥️ #jcdl2026

Their process for associating context with URLs in scholarly docs: 1. Segment to paper components. 2. ID bottom of page 3. Get bib entries 4. ID boilerplate Their annotation was 97.4% reliable every diagreement was reviewed and any confirmed errors were corrected. #jcdl2026

Their baseline and evaluation first required that the URL matching exactly after normalizations. The ref model oerformed the worst vs VisURLX variants. #jcdl2026

"If a URL is in a footnote, you first have to restore it to the sentence where it is used." There are benchmarks for scholarly URL identif. To address the gap, they created a benchmark, created an LLM, compared compression, and evaluated the extraction process. #jcdl2026

Lamia Salsabil (@webscidl.bsky.social 🐸) begins her #jcdl2026 presentation, "End-to-End Contextualized URK Extraction from Scholarly Papers" saying there hasn't been a systematic study for robust methods for extracting URLs w/ their context.

Non-destructive book scanning @ #jcdl2026

Yinlin Chen from @virginiatech.bsky.social descs the library's efforts to digitize photographs at a rate that can't keep up with them being cataloged. #jcdl2026 Off-the-shelf vis-lang models recover less than 10% of the curator's tags. Models can describe the photos but not how curators indexed it.

Fine-tuning helps ~7x but remaining gal is coverage, not ordering. Their pipeline has 3 stages, each targeting coverage, no extra labels. Shuffling the tag order teaches the set. Voting gives a modenfor precision-recall. Many "misses" look like valid subjects that the record left out. #jcdl2026

Next, Ying-Chen Lin from Univ. Missouri talk about their work for identifying valid timestamps for RAG and #webarchiving #jcdl2026 Ads or related stories can have related stamps, too, which are then temporally aligned with the encapsulating page. doi.org/10.1145/3805...

Although in some domains, it doesn't have the highest perf but is better at getting the correct timestamp. Neural modules often got wrong timestamps while symbolic modules often chose the latest stamp. By assigning neural timestamp on high confidence and sumbolic on low -> best perf. #jcdl2026

Boilerplate removal improves input. Neural and symbolic were developed to complement errors. #jcdl2026 Code: github.com/ylbmq/NSTS

At the Historical, Web, and Multimedia Archives session @ #jcdl2026, Bohan Jiang (@brandeisuniversity.bsky.social) descs his archival workflow with CLAMS to find a part of vintage video broadcasts. Using agreement-based selection, they form the segments via 4 Pathways. 📄 doi.org/10.1145/3805...

Normally a good segment is 15-16 sec. long. With the same transcript, all pipelines summarize equally well. #jcdl2026 Future work: indexing summaries and merging and splitting rules that steer segments toward rundown-item length.

Their system includes accuracy of date attribution across domains. They also did an ablation study using only neural, only symbolic, then a combo. #jcdl2026 Using baselines as part of the modules, they can see that standalone symbolic has the best result across domains but combo has higher perf.

Different extraction methods identify and extract date candidates including consideration of confidence and date conflicts from their methods. If confidence, they select a neural timestamp. If low, they select the symbolic timestamp. #jcdl2026

How do you find the boundaries of segments in vintage videos? Through multiple evaluations, including LLMs as judge, they calcuate a median form their scores. They first evaluate native text from the 2nd and 4th pathways. Each is scored in 1-5 against the gold description. #jcdl2026

To start the second day of #JCDL2026, Leif Azzopardi from @microsoft.com is presenting his keynote, "Design for Information Farmers". Leif asks, "How can we reimagine designing for information farmers instead of information seekers?"

It's not clicks and downloads, it's whether the knowledge was sound. Doc isn't the endpoint anymore. Library's competition isn't another repo, it's that the agent is the users' primary interface. #jcdl2026

Leif closes his #jcdl2026 keynote with, "Libraries are the seed repository. Will we still be holding the seed when the next season comes?"

The problem isn't about price, it's about consent. It's both a danger and leverage. Every usage metric will tell that the valuable library is being unused. Agents need curated corpora to give them provenance to trust what they've produced. #jcdl2026

Value is to whoever runs the agent. You're still writing papers, you just don't own the "field" anymore. Nobody likes landlords. Adoption's a race ahead while the controls lag behind. There'll be more to delegate & access where you'll have your own agent, it'll use your DL subscription. #jcdl2026

In a survey, over 90% of attacks were attributed to aggressive bots. Humans vs bot is not the real design problem. Are the bots doing anything for the user or just bull training, exploiting? The new user is the agent. "I probably have more bot users on Twitter than humans." #JCDL2026

Subscriptions must evolve from reading rights to delegation rights. Stop pretending that the agent is human. #JCDL2026

Farming doesn't replace foraging. New varieties farmers plant come from outside the field. We can think about farming as the co-evolutionary loop. Now we have more control over the environment. People interact with information through this metaphor. We're moving from the doc to the task. #JCDL2026

Information farming library isn't categorically different. It can tell when knowledge artifacts become trustworthy. Specific aim isn't individual learning but the cultivation of public ideas. Pruning and weeding become cheaper. The cultivation and knowledge now is much more reliable. #jcdl2026

Foragers themselves were exposed to what was in the wild but farmers worked with a controlled plot. When you build a farm, you have to control for pests (etc). Foraging is linear, with farming you can increase yields, you have higher precision and control about the objs we want. #JCDL2026

At #JCDL2026, "Publishers are connecting agents to their DLs. It isn't the model. It isn't the interface. It's what they sell." It's not the download or document view anymore. We need to be measuring the harvest. They value what comes at the end of the season. The task is the new unit.

Users aren't just seeking information anymore, they're generating it and it's immediately available. People are moving from patch to patch now & growing info. LLMs are a rich mediums to plant a seed (a prompt/RQ) then you grow out from this info. In the end: you harvest what you created. #JCDL2026

At the first day of #JCDL2026, Nicholas Belkin (@rutgersuniversity.bsky.social) starts his keynote "Digital Libraries and Their Technologies: A Look Back and a Peek Forward" by reminiscing on his involvement in organizing the conference 25 years ago under the auspices of @tamu.bsky.social.

LLMs have met many of the proposals of Libraries of the Future. How well do early ideas of LotF address claims of what libs should do. The push to focus on a particular kind of community, knowledge, and info missed the boat. They're addressing what serving a community of users might be. #JCDL2026

Provenance and protection of personal property? LLMs do not meet this requirement of LotF. There are things that libraries do that LLMs don't do that are the functions of a library. e.g., maintaining history, quality control. Libraries support communities. LLMs do not. -30- #JCDL2026

In 2020s, retrieval was greatly increased through vector embedding & not just language models but LLMs. These advances have been very significant to the dev of DLs. They've been applied to almost all functions of the DL & have been used to support auto classification, retrieval, etc. #JCDL2026

Have Digital Libraries met the requirements for the proposal Libraries of the Future? • Universal remote access to the store of all knowledge? ✔️ • Linking of collections? ✔️ • Direct access to the core of knowledge? ✔️ • Providing at least 2 levels of representation? ✔ #JCDL2026

There was a broad range of interest addressing a broad range of functions in various ways w/ lots of different techs. This Phase 2 of DL creation facilitated the use of DL in education – DLs for a purpose & understanding DLs as libraries (which DLs are) as central to the edu enterprise. #JCDL2026

Do LLMs fulfill all of the proposal details for libraries of the future? Is this the end of DL practice as we know it? DLs have successfully applied techs to this functions through the construction of individual libraries and #webarchiving. #JCDL2026

Sharing of content, even metadata, was a big problem & something that needed to be addressed in the early days of DLs. New techs like CD-ROMs encouraged wide-scale sharing of content. They were put into use to access store of knowledge, breadth of cultural heritage, and depth of access. #JCDL2026

Does anyone remember Gopher? ✋ These changes in tech linked significant passages and activities across documents. The TREC program looked to access to large amounts (~ 1 GB at the time) of machine-readable docs. #JCDL2026

NSF sponsored a bunch of informal workshops in the late 80s / early 90s in response to what was going on in the digital libraries field. Early DL meeting history began w/ invitational workshops. In '94/'95 were the 1st and 2nd annual confs on the Theory & Practice on Digital Libraries #JCDL2026

Unsurprisingly, people elsewhere in the world beyond the US were interested in digital libraries. The Libraries in the Digital Age initiative was to discuss not just digital libraries but what libraries could potentially become. #JCDL2026

Common elements of these Libraries of the Future proposals: • universal, remote access to the store of human knowledge • at least two levels of knowledge, not just the information • meaningful linking of docs to one another and the constructed representation #JCDL2026

What are the primary social functions of libraries? #JCDL2026

In describing Licklider's Procognitive Systems, Nicholas recounts their definition for libraries of the future, which didn't involve any librarians so as to not be constrained. "If we're going to talk about libraries of the future, we're not talking of libraries now but something else." #JCDL2026

Libraries should facilitate their own development and facilitate the joint contribution of users to the body of knowledge and the use of the body of knowledge. Information retrieval should change its goal from dispensing information to understanding information. #JCDL2026