OpenAlex keywords used to be bad. Now they’re good. 419 million works in OpenAlex have new keywords, drawn from a vocabulary of 1.88 million. They’re 91% accurate, up from 44% for the old keywords.

The old keywords were guessed from each work’s topics, so they were often off. The paper announcing the first gravitational waves got “Black hole (networking)” and “Binary number.” The new tagger reads each work’s title and abstract and says what it’s about:

Paper	Old keywords	New keywords
Observation of gravitational waves from a binary black hole merger (2016, here)	Black hole (networking), Binary number, Computer science	binary black hole mergers, LIGO, ringdown, tests of general relativity
Deep learning (2015, here)	Representation (politics), Layer (electronics)	backpropagation, convolutional neural networks, recurrent neural networks
Highly accurate protein structure prediction with AlphaFold (2021, here)	Computer science, Function (biology)	AlphaFold, protein structure prediction, protein folding problem, multi-sequence alignment

How much better

Our judge is an AI model from another lab: OpenAI’s GPT-6 Astra, since our keywords were built with Anthropic’s Claude. It read 400 random works and rated each keyword on its own. You can see the results above. Shown both lists side by side, the judge preferred the new keywords for 395 of the 400 works.

The new keywords also match the keywords authors chose for their own papers twice as often as the old ones did, and they beat our previous best models on two standard keyphrase benchmarks. Every number is here.

The new keywords cover 89% of works, down from 95%. That’s on purpose: when a record doesn’t say what it’s about (a bare dataset entry, a table of contents), the tagger now returns nothing rather than a guess.

What they’re good for

• Finding what text search misses. A keyword finds the paper even when authors used different words for the same concepts. 133 million works with no abstract still get keywords.
• Mapping a field. Group any set of works by keyword to see what it’s made of, and by year to see what’s rising.
• Finding experts. Filter to a keyword and group by author; in our tests the top ten were real specialists on 24 of 24 queries.
• Searching across languages. Keywords are in English, but papers in any language get them: a Chinese paper on microplastics in soil carries the same microplastics keyword as an English one. About half of what the keywords add to a search on microplastics or remote work isn’t in English.

How to: find the papers your text search misses

Say you’re studying antimicrobial resistance. Just search: OpenAlex search now reads keywords along with titles and abstracts. When a phrase you type names a keyword, a paper matches if the phrase is in its title or abstract or the paper carries that keyword. A title and abstract search for the phrase finds 118,920 works; this one finds 241,167:

https://api.openalex.org/works?search.title_abstract_keywords="antimicrobial resistance"

The antimicrobial-resistance keyword brings in about 122,000 of them, and our judge rates 81% of those on topic: about 100,000 relevant works the phrase search missed, so you find about twice as many. Some say “antibiotic resistance” or “drug-resistant bacteria” instead. 47,269 have no abstract at all, and 74% of those are on topic, like “Increased resistance to antibiotics among microorganisms isolated in a general hospital” (1958, here) and “Lack of development of new antimicrobial drugs” (2005, here); no text search could find those.

On openalex.org there’s nothing to set: “Title, abstract, & keywords” is now the search box’s default, and the top results are reranked so the papers most clearly about your topic come first. Every other word you type still has to be in the text, so “remote work productivity” finds papers tagged remote work that mention productivity, never a paper tagged only productivity. In OQL it’s works where title-abstract-keywords has ("antimicrobial resistance").

Let your AI agent do a thorough search

The search catches the keywords your own words name. An AI agent can go further: find the keywords that mean your topic in other words, add them, and read the results to drop the strays. The /text/keywords endpoint takes any text, including a title or abstract, and returns the keywords that fit it, and an agent like Claude Code or Codex can use it for you. Tell it something like this:

Check help.openalex.org first, then use the OpenAlex API to find papers on <your topic>. Use the title, abstract and keywords search, and add OpenAlex keywords that describe my topic in other words, all in one query. Each paper should cover every part of my topic. Read the results as you go and keep only the papers that are really about my topic. Save the 200 most-cited of those as a CSV and show me the query and how many papers each part found.
My API key: <your key>

Keep the “check help.openalex.org” part: without it, agents work from what they remember about the OpenAlex API, which is out of date. We tested this prompt on fresh installs of Claude Code and Codex: every run found the title, abstract and keywords search on the help site and built one query with it, and on remote work and employee wellbeing, 99% of the papers in their final lists were at least partly on topic (about seven in ten clearly so). In the Claude app, set the chat to Auto (next to the model name, under the message box), or it will ask your permission before every search. You’ll need a free API key; the full recipe is here.

Or, if you use Claude, just ask. Add the free OpenAlex connector, now in Claude’s connector directory, and ask in one sentence: “Do a thorough search for open-access papers on how remote work affects employee wellbeing.” The connector searches titles, abstracts and keywords, adds keywords that mean your topic in other words, reranks the results so the most relevant come first, tells you how many works each part found, and hands back the query so you can rerun it.

The vocabulary keeps changing

This isn’t a fixed list. New fields show up every year, and we’ll add keywords for them as they do. We already merge keywords that mean the same thing (“Neanderthals” and “Neandertals”), and we’ll keep merging, splitting and adding as you tell us what’s wrong. Send us feedback on any keyword that misses.

Keywords are already linked to Wikidata: about 22% of keywords have a Wikidata id, and those cover about 56% of keyword uses. That makes it easy to join OpenAlex keywords to other data. Keyword search also knows 5 million synonyms, so searching keywords for “heart attack” finds myocardial infarction. Next, we’ll look at joining keywords to other outside vocabularies like MeSH, and at vector search, so that searching the keywords for “antibacterial resistance” finds antimicrobial resistance. We also want to assign each keyword to one or more of our topics, so keywords become part of the whole OpenAlex aboutness taxonomy.

Check our work

Everything is open: the code, prompts, benchmarks, vocabulary and model weights, with a release that lets you rerun the tagger yourself. The keywords help page has the details.

Edit, 3 October 2026: An earlier version of this post showed how to run a title and abstract search and a keyword search side by side and combine them by hand. A few hours after it went out, we shipped a search that does both at once (title, abstract and keywords, now the default on openalex.org), so we rewrote the how-to section and the AI prompt to use it.

Edit, 3 October 2026, later the same day: The OQL field in this post is now spelled title-abstract-keywords (it was title/abstract/keywords), because slashes already separate the parts of OpenAlex IDs. The old spelling still works.

You can now filter all of OpenAlex by study design: randomized controlled trials, systematic reviews, meta-analyses and more. 21 million works, in every field and language, carry one. Our randomized-trial tags are 99.7% precise, and on the same papers they’re more precise than PubMed’s, long the gold standard.

When you’re searching, you can focus on just the highest-quality evidence. And if you build evidence tools, you get study design for the whole literature, free, through the OpenAlex API.

Seven designs, every field

The designs come from PubMed’s own vocabulary, and we add nothing to it. What’s new is reach. PubMed tags the journals it indexes; we tag every work with an abstract. Half of the randomized trials we tag, 306,000, are in journals PubMed’s MEDLINE doesn’t index.

Here’s a graph how many works are tagged with each study design, in both OpenAlex and PubMed:

Every randomized trial is also a clinical trial, and every meta-analysis is also a systematic review, so a work can count twice. PubMed’s Observational Study tag dates only from 2014. Works without an abstract get no design.

Find the trials, see the evidence

Filter to one design. Of 34,395 papers that mention semaglutide, 533 are randomized trials; one filter returns just those, on openalex.org or in the API with filter=study_designs.id:randomized-controlled-trial. AI agents that search OpenAlex can do the same.

Group by design to see what kind of evidence a field has. Long COVID research is mostly observational: 31 observational studies for every randomized trial. Semaglutide already has more meta-analyses (1,341) than randomized trials.

Tagged papers on each topic, by design. Papers an OpenAlex search for the topic returns.

Share of each topic’s papers that carry a design; protocols left out. We tag conservatively, so real counts run higher, trials most of all.

Built on PubMed, tuned for precision

PubMed’s publication types are carefully indexed and trusted across medicine. But its tags sometimes describe the study a paper came from rather than the paper itself, so a methods paper or a later analysis of trial data can carry the trial’s tag. We tag only what the paper itself reports, and only when we’re confident.

How often each tag is right

Share of tags an independent judge confirms, on the same PubMed-indexed papers.

Where the tags differ

Here are some examples of papers that PubMed tags as RCTs, but we (correctly) don’t apply the RCT.

Paper	What it is
Bland and Altman, Lancet 1986: measuring agreement	A statistics methods paper
TOAST stroke classification, 1993	Definitions written for a trial
The Implicit Association Test, 1998	Lab experiments that built a test
Oncotype DX gene test, NEJM 2004	A gene test on tumors from one arm of an earlier trial
SAPS II severity score, JAMA 1993	A risk score built from a cohort

Our model is tuned for precision, not recall–when we say it’s an RCT, it’s an RCT, with 99.9% accuracy. PubMed captures some RCTs that our stingy approach misses, but at the cost of some accuracy.

Judge: Claude Opus 5.5, under written definitions, on 8,308 papers sampled so that every estimate is weighted to the whole index; nothing was tuned on that sample. Re-reading 436 open-access papers in full barely moved the verdicts.

Check our work

Everything behind these numbers is public: the code, the models, the samples and every judge’s verdict. One command reruns every table. github.com/ourresearch/openalex-study-designs

When people want to know whether a piece of research is any good, they usually have two options: ask an expert, or count citations. Both have well-known problems. Expert review is slow, expensive, subjective, and does not scale to portfolios of thousands of papers. Citations take years to accumulate, reward the already well-cited, and say more about whether a paper was used than whether it said anything new.

That second gap is the interesting one. Creative work, including scientific discovery, is generally understood to be both new and useful. Citations are a reasonable proxy for usefulness. We have had no widely trusted, scalable measure for new. Funders say they want novel research, and then evaluate it with tools that cannot see novelty.

Earlier this year an international challenge set out to change that, and the winning approach now exists as a working prototype. We are considering adding a novelty score to OpenAlex. Before we write a spec or scope resources to implement it, we want to know whether you would use it, what for, and what would make you trust it.

The challenge

The Metascience Novelty Indicators Challenge was launched in September 2025 by the UK Metascience Unit at UK Research and Innovation; designed by the Science Policy Research Unit (SPRU) at the University of Sussex and RAND Europe; and operated by Challenge Works. Teams were asked to build a tool that could assess the novelty of a research paper at the time of publication.

To evaluate the entries, SPRU built a ground-truth dataset the hard way: a random sample of recent publications across all disciplines (drawn from OpenAlex), each rated for novelty by human experts in the relevant field who were provided with the full text. Challenge entrants never saw those ratings. Their indicators were scored on how closely they matched the experts’, and the evaluators were blinded to who had submitted what.

The winner, announced just a few months ago in June 2026, was a team from Forschungszentrum Jülich in Germany: Jan Göpfert, Samuel Kieling, Jann Weinand, Titan Hartono, and Patrick Kuckertz. Their approach came top on most evaluation criteria (in a “winner of winners approach”), and they received a £300,000 prize, provided by Coefficient Giving, to develop it further.

Coverage: https://www.nature.com/articles/d41586-025-01882-7 and https://www.science.org/content/article/how-novel-research-paper-competition-quantify-concept-crowns-winner.

How the Jülich indicator works

Most existing novelty indicators work from metadata: unusual combinations of references, keywords, or journals. The Jülich indicator reads the paper instead.

It uses a large language model to analyse the study and a selection of the works it cites, and from those, it reconstructs the state of knowledge at the time of publication, including the open questions in the field. It then asks what the focal study contributes anew: an original method, a surprising result, or a solution to a problem that was previously unsolved. It deliberately collects arguments for and against the paper’s novelty and weighs them. The output is a score from 0 to 100, an interval showing how confident the model is, and a written justification that a reader can check.

Jülich press release: https://www.fz-juelich.de/en/news/archive/press-release/2026/ai-identifies-scientific-novelty-julich-team-wins-international-challenge

The team is clear that this is a support for human judgement rather than a replacement for it, and that a score which could be gamed, or which quietly disadvantaged some fields or languages, would be worse than no score at all. Those are our concerns too.

Why OpenAlex is interested, and what it would take

OpenAlex already carries citation counts, topics, and a growing set of work-level signals (e.g., field-weighted citation impact). A novelty score that is transparent, reproducible, and available for every work, free, in the same place, would be a genuinely new thing for funders, institutions, journals, and researchers to work with. The UK Metascience Unit has told us that it would be of interest to a large proportion of the roughly 100 research teams it funds, as well as to UKRI’s own analysts. And we would rather that such a novelty indicator be built in the open than watch it appear behind a paywall.

It is not a small undertaking, and the open questions are the reason for this post:

• How far back? Scoring every work in OpenAlex is very different from scoring everything published since, say, 2019.
• Which works? Journal articles only, or preprints, conference papers, book chapters, and dissertations as well? The indicator was validated so far only on journal articles. Maybe that’s enough (after all, FWCI only applies to a few work types), but maybe it’s worth the effort to evaluate the indicator in a broader set of work types first. 
• One-off or continuous? A static file scored once is much cheaper than a live score that updates as new works arrive and as records change. A live production database like OpenAlex really wants the latter, but maybe the community only needs static scores, updated yearly.
• Open-weight or commercial models? Open-weight LLMs would make the scores reproducible over the long term and could run on public research computing. Jülich is testing whether they perform as well.
• Cost and compute. Running a language model over hundreds of millions of works is a significant compute bill, however it is done. That’s why answers to the questions above are especially important before we fully spec this out.
• Transparency versus gaming. Publishing the full method makes the score checkable and makes it easier to write to. Where should the line sit and is that an argument for static historic application vs. on-going scoring of the indicator against new works?
• Use in evaluation. A score built to describe a literature can be misused to judge individuals. What safeguards would you want before it existed at all?

What we are asking

This is a vibe check with a light-touch questionnaire, not a full consultation. We want enough signal on demand and scope to write a spec and a resourcing case, and we will go deeper later if the response warrants it.

Fill in the form. About six questions, five minutes, any language. It stays open until Friday 13 November 2026.

https://docs.google.com/forms/d/e/1FAIpQLScOlmbDN7AdJic8jHO1dnQWMhppFIrMK96J0s6TmcU1oME3ZQ/viewform?usp=publish-editor

Join the webinar. Wednesday 7 October 2026, 8:00 to 9:00am Pacific (11am Eastern, 4pm UK, 5pm Central Europe). Ben Steyn (UK Metascience Unit) on why the challenge was run, Jan Göpfert (Jülich) on how the indicator works, and Kyle Demes (OpenAlex) on what integrating it into OpenAlex would look like, followed by half an hour of questions. Sarah Otner (SPRU) will be on hand for questions about the evaluation. Recording and slides posted afterwards.

https://zoom.us/webinar/register/WN_EetlA-l3QgK2FvQVxE0GLg

Pass it on. Funders, research offices, metascience groups, tool builders, and anyone who has ever wished that citations told them something they do not.

What happens next

We will share a synthesis of the responses with the UK Metascience Unit and the Jülich team in late November, and publish a summary here. If there is clear demand and a workable scope, the next step is a spec and a scoped first run. If there is not, we will say so and keep the door open.

Thanks in advance for any input you share. If you have any other questions about this initiative, let me know at: kyle@openalex.org 

Since 2018, Unpaywall and then OpenAlex have described the open access status of a work with a colour: gold, green, hybrid, bronze, or closed. The labels were a pragmatic answer to a practical question (“can I read this, and where?”) and they caught on. Today they are embedded in major bibliometric databases, funder compliance dashboards, institutional reports, and national open access monitors around the world.

They have also started to crack. Earlier this year the OpenAlex Community Advisory Board convened a working group to look at the problems and recommend a way forward. The group, chaired by Juan Pablo Alperin, published its report last month:

Alperin, J.P. et al. (2026). Recommendations for Revising the Open Access Classification System in OpenAlex. Zenodo. https://doi.org/10.5281/zenodo.21934115

Before we further consider such significant changes to the open access classification in OpenAlex (and Unpaywall), we want to know whether the problems the working group identified match your experience, and whether the direction it proposes makes sense for the way you use open access data. This post explains both, and asks for ten minutes of your time.

What has gone wrong with the colours

The working group’s diagnosis comes down to four things.

• The colours were built for journal articles. Preprints get called green, which implies they are a copy of a publisher version that may never exist. Dissertations live in university repositories and fit nowhere. Datasets and software are rarely “published” by a publisher at all. Books can have open chapters and closed chapters in the same volume. None of this maps onto a scheme that assumes every work has a publisher version and possibly a repository copy.
• They lean on journal-level policy instead of what is true of the work. Gold is assigned because a journal is listed in DOAJ or is empirically mostly open, not because the article in front of you is open. Journals flip, get acquired, adopt subscribe-to-open, or publishers change their mind. A journal can be gold today while all of its back catalogue is closed. This also cuts against the direction the research assessment community has taken with DORA, which asks us to look at the work rather than the venue.
• They mix up access with business model. Gold tells you something about how a journal is funded and something about whether you can read the article, and it is not always clear which. Green does not say whether the repository copy carries a licence you can reuse. Bronze is a catch-all for “free to read, no idea why”, and a work can move in and out of bronze several times in a year.
• They have no sense of time. An embargoed article that opens after twelve months, or a free article that a publisher later closes, gets a single snapshot label. Anyone tracking compliance or studying how access changes over time has to work around that.

What the working group proposes

Instead of one label that tries to say everything, the report recommends describing each work along three separate dimensions:

• Access: is the work free to read, or closed?
• Location: is the free copy at the publisher, or on another platform such as a repository or preprint server?
• Licence: does the work carry an open licence, a restrictive one, or none that can be identified?

Combined, these give plain-language, work-level labels such as Publisher Open, Platform Public, or Closed that work the same way for an article, a preprint, a dataset, or a book chapter. Venue-level questions (is this an open access journal?) are left to be derived from the works themselves, or handled by a separate classification.

Just as important is what the proposal does not try to do. It says nothing about business models: APCs, diamond, hybrid agreements, subscribe-to-open, transformative agreements. The working group’s view is that business model is a property of a venue, access is a property of a work, and squashing them into one label is where much of the confusion came from. They recommend a separate effort on a venue business model taxonomy, and we agree.

The report includes a mapping from the proposed labels back to the current colours, use cases from funders, institutions, publishers, and database providers, and a discussion of transition. It is 15 pages, summarizing deliberations from thought-leaders in the space and worth reading in full.

What we are asking

This is a vibe check, not a line-by-line review. We want to know three things: does the problem resonate, does the three-dimension direction make sense, and what would you need from us to adopt it?

Fill in the form. About six questions, five minutes, any language. It stays open until Friday 13 November 2026.

https://docs.google.com/forms/d/e/1FAIpQLSew0id7AefKryr-SGeb8DChSQkxoMUM6WP2ebmV7U3hVjuo3A/viewform?usp=publish-editor

Join the webinar. Tuesday 6 October 2026, 7:00 to 8:00am Pacific (10am Eastern, 3pm UK, 4pm Central Europe). Juan Pablo Alperin will present the report, and Ashley Farley (Gates Foundation, working group member) and I will join for Q&A. The recording and slides will be posted afterwards.

https://zoom.us/webinar/register/WN_NbIKwVu6Q4SI83Fv5qwXxw

Pass it on. If you know people whose reporting, policies, or reimbursement workflows depend on the colours today, please send this to them. Those are exactly the voices we most need to hear before anything changes.

What happens next

Nothing changes in the OpenAlex API or website as a result of this post. We will synthesise the responses and take them to the Community Advisory Board on 10 December 2026. If the community response supports the direction, the next phase would be implementation planning: a transition period with old and new labels available side by side, direct engagement with the databases and tools that reuse our open access data, and a separate conversation about business models. If the community does not support the proposal, we will have learned that cheaply, and we will say so.

Thanks to the working group for the report: Juan Pablo Alperin, Amélie Church, Nees Jan van Eck, Ashley Farley, Rupert Gatti, Melissa Harrison, Bianca Kramer, Christopher Marcum, Dom Mitchell, and Roberta Takenaka. 

If you have any questions about this initiative more generally, send me an email at kyle@openalex.org 

Researchfish is closing on 31 July 2027. For the more than 90 funders that relied on it, including UKRI, NIHR, and many of the UK’s largest research charities, the question of how to track what their funding produced is suddenly urgent. When UKRI last went to tender, it concluded there was no viable alternative product (Johnson 2026). We think that was the wrong question. The alternative to Researchfish is not another product. It is open infrastructure that is already carrying this data at scale, plus a policy decision about where researchers assert the link between their grants and their outputs.

In this post we (Melissa Harrison, Europe PMC and Kyle Demes, OpenAlex) want to show what already exists, what our two organisations are connecting next, and what funders can do today.

Two different jobs, only one of which belongs in the open

Progress reporting bundles together two very different kinds of information, only one of which belongs in the open, so it is important to be precise about which one we are talking about. Narrative reporting, such as progress against objectives, team composition, and unpublished or sensitive results, belongs in funders’ own systems, and nothing in this post suggests moving it.

Output reporting is different. Which publications, preprints, datasets, software, and clinical trials resulted from which award is factual information that already lives in the scholarly record. Yet for years, researchers have been asked to retype it annually into a proprietary platform, where it was locked away from the very community that produced it. That is the part of Researchfish’s job that open infrastructure can take over, and the part this post is about.

This is not hypothetical. It is running today.

The infrastructure exists

Europe PMC and OpenAlex play distinct but complementary roles, and the pipeline between them is already operating at scale.

Europe PMC is a world-leading open biomedical literature resource that combines a discovery platform with a full-text repository. Through its submission system, Europe PMC plus, researchers funded by 35 Europe PMC funders already link their publications directly to the grants that supported them, at the moment they deposit. Those grant–publication links live in Europe PMC’s open database connected to other core resources like the life science datasets underpinning these publications.

OpenAlex is the open scholarly graph: hundreds of millions of works connected to authors, institutions, sources, and funders, all openly licensed with open APIs. OpenAlex now holds over 16 million award records from roughly 30,000 funders worldwide, drawn from funders’ own open databases (Gateway to Research, NIH ExPORTER, Wellcome’s open grants data, and dozens more) and from assertions in the scholarly record. More than half of those awards are already linked to at least one research output, and those outputs are not just journal articles: they include hundreds of thousands of datasets, preprints, reports and dissertations, as well as software and more.

This is the part that makes this a report on progress rather than a proposal: Europe PMC is already one of the largest sources of grant–output linkages in OpenAlex, with millions of researcher- and funder-connected assertions flowing from Europe PMC into the open graph. A researcher who links a paper to a grant at deposit in Europe PMC is already contributing to an open, reusable record of what that funding produced.

Publisher-supplied funding metadata (the kind that flows through Crossref and DataCite) has incredible value,  however, researchers have no easy way to correct or extend it after publication and the process relies on publishers delivering this metadata when registering DOIs. Europe PMC and OpenAlex are the places in the open ecosystem where researchers themselves can assert and confirm grant-output links. That researcher-confirmed layer is precisely the human-curated data Researchfish was built to collect. It already exists in the open, it just needs to become the default place this reporting happens.

Open means many front doors

A proprietary platform gives every funder the same front door and a single supplier. Open infrastructure works differently, and this flexibility is its core advantage. A funder deciding what to do before July 2027 has several routes, all available now:

Use Europe PMC as the reporting platform. For existing Europe PMC funders, grant linking at deposit is in production today. Point grantees to a workflow they already use for open access compliance, and the linkages accumulate continuously in the open instead of arriving in an annual burst. For other life sciences funders, becoming a Funder of Europe PMC is the closest thing to a like-for-like replacement for the outputs portion of Researchfish, and it exists now.

Work with OpenAlex directly. Funders across all disciplines can have their awards represented in OpenAlex and build reporting and analysis straight on the open graph, covering output types and fields well beyond biomedicine. Many funders are already doing this.

Integrate open infrastructure into systems you already run. A funder can keep its grants management system or portal and pull candidate outputs from open APIs, asking researchers only to confirm a funder or grant ID rather than to type it. One funder OpenAlex works with already compares open grant-output linkages against its own compliance records and automatically reminds exactly the researchers whose assertions are missing or their requirement. Reporting becomes exception handling instead of data entry.

Both Europe PMC and OpenAlex provide open APIs and openly licensed data, so research information system vendors and other open infrastructures can build on the same foundation. With a proprietary platform, you choose a supplier. With open infrastructure, you choose a workflow and you can change your mind later without losing the data, because the data was never trapped inside a platform.

What OpenAlex and Europe PMC are working on together

The foundations exist; our collaboration is about tightening the loop between them. The directions we are exploring include making the flow of researcher assertions from Europe PMC into OpenAlex fully systematic and timely, so a link confirmed at deposit is visible to funders on OpenAlex within days; presenting researchers with candidate linkages from text mining and the open graph to confirm rather than enter from scratch; and capturing links to datasets, software, and other outputs at the same natural moment in the Europe PMC workflow. We will share more as this work develops, and we would rather shape it with the community than announce it finished.

What funders can do now

Two decisions, both available today, would do more than any new platform:

Publish your award list openly. Even a minimal feed of award IDs and titles, without personal data, gives researchers something to select instead of retype, lets open infrastructure link outputs to your grants, and means your reporting data accrues in the open from day one. Funders like Wellcome and UKRI (through Gateway to Research) already show this is routine.

Require funding metadata at the point of deposit, and accept researchers’ assertions in open infrastructures as reporting. When linking the funder and grant is part of the process of depositing the output, the information is captured once, at the natural moment, and can be reused everywhere. The annual reporting burst simply stops being necessary.

Join the conversation

Researchfish’s closure sets a deadline, but it is also the first real opening in a decade to move output reporting onto infrastructure the community owns. If you are a funder weighing your options, an institution with reporting requirements, or an open infrastructure interested in connecting, we would love to hear from you: 

Melissa Harrison mharrison@ebi.ac.uk 

Kyle Demes kyle@openalex.org

Research relies on software. Software written by scientists, for science, runs through the entire modern research stack: NumPy and SciPy, R and ggplot2, Jupyter, BLAST, ImageJ, AlphaFold. Yet in the scholarly record, that software is nearly invisible. Software is not usually cited formally in publications and is usually just mentioned in the text, which means the software (and the people who build and maintain it) never shows up in the databases the research world uses to find, measure, and reward scholarship.

We’re going to fix that. Thanks to a 2-year grant from Schmidt Sciences, OpenAlex is joining forces with a team at University of Texas at Austin led by James Howison to make research software a first-class part of the open scholarly graph. It’s a problem we’ve been chipping away at for a long time: over a decade ago we built Depsy as an early proof of concept for measuring the impact of research software. Now we get to do it properly, at the scale of the full scholarly graph. Work kicks off in mid-August 2026 (you can read the full proposal here).

Why software, and why now

The scholarly graph records the relationships among papers, authors, institutions, funders, and topics. It’s how researchers find work, how contributors get credit, and how funders and institutions make decisions. But because software so rarely appears in reference lists, it can’t be found or measured in any of today’s scholarly databases, open or proprietary.

That invisibility has real costs. Research software engineers whose tools underpin entire disciplines struggle in promotion and tenure processes because their contributions don’t “count.” Funders can’t see where software investments would make the biggest difference, or what impact past investments have had. And as AI coding assistants become part of everyday research practice, the missing link between papers and the code behind them is exactly the context those systems need most.

What you can expect in OpenAlex

This grant funds a set of major additions to OpenAlex that users have been asking about for years:

Software as a research entity. We’ll seed OpenAlex with an initial registry of software projects and add software as a work type (in fact, we just launched this), connected to the rest of the graph: authors, institutions, topics, funders, and the papers that use it. Each software entity will link out to external identifiers (package registries, repository URLs, DOIs, Software Heritage IDs) with full provenance.

Software mentions, extracted from full text. Most software never makes it into reference lists, so we’ll find it where it actually lives: in the text of papers. We’re building a mention-extraction and disambiguation pipeline that links papers to the specific software they use, and we’ll expose the results through a new citation object, complete with the surrounding text context, the citation type, and confidence scores, all searchable through its own API endpoint.

True versioning, for software and for all works. Software changes constantly, and “which version did they use?” matters for everything from reproducibility to impact measurement. We’re adding a generalized versioning model to the OpenAlex schema. Software is the first beneficiary, but versioning will extend to other work types where it’s increasingly important, like datasets and preprints.

Credit for the people who build research software. Author profiles will gain software-specific contribution and adoption metrics, so the maintainers of foundational tools can finally show where their work is used, much as in  systems that already track publications.

Along the way, we’ll be upgrading the machinery underneath: broader full-text coverage, faster PDF parsing, and improved author disambiguation, all of which will improve OpenAlex well beyond software.

All of it will be open, as always: CC0 data, open source code, and open documentation, available through the OpenAlex API, UI, and snapshot.

We’re hiring

OpenAlex will build and operate the production systems, while a team at the University of Texas at Austin, led by James Howison, brings deep expertise in research software scholarship and software mention extraction. A project this size needs more hands. OpenAlex will be opening several new positions in the next few weeks, spanning data and software engineering and community management roles. Keep an eye on openalex.org/jobs, and if this is the kind of problem you want to work on, we’d love to hear from you.

Thank you

We’re grateful to Schmidt Sciences for backing this work. Making research software visible has been a goal of the research software community for more than a decade, and we’re honored to help build the open infrastructure that finally makes it routine.

As always, we’d love your feedback: if you have use cases for software metadata in OpenAlex, ideas, or concerns, get in touch at support@openalex.org.

—The OpenAlex Team

*This post was written with Claude. Every number and example was verified by a human against the live data.

Unlike most scholarly databases, OpenAlex doesn’t start from journals. Other databases pick a journal, index it, and attach its articles — so their journal metadata is clean by construction. OpenAlex works the other way: we index scholarly works first, then connect each one to the rest of the research graph — authors, institutions, and publication venues. Venue information arrives from many different sources, unstandardized and often without persistent identifiers. That’s how one journal can quietly become two or three records. And because source metadata is critical for many of our users — especially librarians — it was time for a comprehensive cleanup.

Over the past two weeks we merged 27,728 duplicate source records — the same journal listed under a translated name, an old name, or a small spelling difference, its papers split among the copies — correcting the venue records of more than 6 million works. After the cleanup, those papers live under one journal: one complete works count, one citation profile, one page. The active catalog went from 282,924 source records to 255,434; every removed record was a duplicate, and no real venue lost its page.

Three examples

Chemischer Informationsdienst → ChemInform. Wiley renamed this chemistry alerting service decades ago — but the catalog carried both names as separate journals, splitting fifty years of chemistry down the middle: 307,000 works under the German name, the rest under the English one. It’s one record now, 794,000 works, with the German identity preserved as an alternate title and searchable in both languages.

Journal of the American Medical Association → JAMA. Even the world’s best-known journals weren’t immune: 85,611 papers were filed under the spelled-out name as if it were a separate journal from JAMA. Roughly a sixth of JAMA’s output was invisible from its own page. One record now — 354,735 works — and the spelled-out name remains searchable as an alternate title.

镇江医学院学报 → Journal of Zhenjiang Medical College. This journal existed twice in the catalog — once under its Chinese name, once under its English one. The telling detail: the duplicate’s publication activity stops in 2001 — the exact year Zhenjiang Medical College merged into Jiangsu University. The bibliographic data echoed institutional history; the two records are now one.

How it works

The hard part of this work isn’t finding lookalike records, it’s that academic publishing is full of journals with similar names that must not be merged. Every merge decision ran through four independent layers:

• Full-evidence judgment, never name matching. Each candidate pair was decided on its complete records: ISSNs, publishers, countries, and year-by-year publication activity. Two journals publishing independently in the same years is evidence of two real journals, no matter how similar the names.
• An adversarial second pass. Every approved merge went to a second model with the opposite job: find concrete evidence the two records are different venues. Only merges that survived their own prosecution proceeded.
• An independent re-check of every consequential pair. All 12,660 merges involving journals with 50+ papers were re-reviewed by a separate model with fresh evidence. This layer caught things the first two missed — for example, a proposed merge of 터널과 지하공간 (the Korean Society for Rock Mechanics’ journal) with a similarly named but distinct tunneling journal carrying different ISSNs. That layer alone caught 570 proposed merges that were wrong.
• Human audits, on random samples. Merges were hand-checked throughout: a stratified 50-pair review sheet was graded before each execution wave, plus a purely random sample drawn from the merges as they shipped — zero were rejected. And the largest merges were never automated at all: every pair whose retiring record held 10,000+ works was individually reviewed and approved by a human, with the four disputed cases held out entirely.

More than 158,000 examined pairs were confirmed distinct — disjoint ISSN families, different publishers, independent publication histories — and recorded, so future passes build on settled ground instead of starting over. This was a first comprehensive pass; as metadata improves, more merges will follow.

What you’ll notice

• Journal pages with more complete works counts and consolidated ISSNs.
• Search that works in every language a journal has ever been named in — merged records keep every historical and cross-language title. One Czech agriculture journal now carries all five names it has published under since the 1950s.
• Retired duplicate IDs now return 404. If you have stored source IDs that stopped resolving, the surviving record has everything — papers, ISSNs, all names — and searching by ISSN or by any of the journal’s names will find it.

What’s next

New records arrive every day, and the pipeline that did this — find candidates, judge on full evidence, verify independently — is now a repeatable tool for keeping the catalog clean. Next up for source metadata: cleaning up the publisher and host-organization lists, and improving how works get matched to their journals in the first place.

TL;DR. Research security screening checks author affiliations against lists of organizations of concern, using bibliometric data. Affiliation matching is imperfect, though, especially for companies and small institutes with short names or no ROR record, so a researcher can end up linked to an organization they have no connection to. As a first pass we cross-referenced the major watch lists to ROR and cleaned up 6,191 such over-merges across 192 listed organizations. We invite the research security community to collaborate with us to improve OpenAlex and use bibliometric data responsibly for this use case.

---

Research security screening is now routine. Funders, universities, and commercial compliance tools check author affiliations against government lists of organizations of concern (e.g., entities flagged for military, defence, or sanctions reasons). The lists are public and the screening is increasingly required.

One part of that screening process often relies on affiliation data from research publications. The general idea is to ask: has the research team been associated with organizations of concern directly (i.e., they worked for or were funded by them) or indirectly (through their networks of collaborators). A positive match in a bibliometric database then triggers follow-up scrutiny to determine if the connection is real and the risks associated with the connection.

Anyone who has worked with affiliation data knows that it can get messy. To demonstrate the problem, the University of British Columbia has almost 200,000 distinct text strings that have been used in publications to refer to the University, covering departments, addresses, translations, typos, text encoding issues from publishers, and more. Here are a few examples:

• “The University of British Columbia”
• “UBC”
• “Univ of British Columbia, Vancouver, BC, Canada”
• “University of Brtish Columbia, Vancouver, Canada”
• “University of British Colombia#TAB#”

And so algorithmic approaches are used to associate these text strings with institutions in the Research Organization Registry. Research information databases each have their own processes for doing this. We make our source code openly available, and you can find our methods detailed on github. But long story short, we have linked over 150M unique text strings to 130k+ unique institutions. These algorithms generally work well, with >98% precision and >90% recall for most academic institutions. But across the full corpus of all scholarly works in OpenAlex (500M+), that’s still millions of errors.

These errors can be especially acute for companies or small research institutes because they may have very few publications to train the matching model and they often have short names that are common words. Consider the example of the company “Optical Sciences”. They have a ROR record because they publish research, but they get erroneously associated with the Optical Sciences Center of the University of Arizona which does not have its own ROR record.

Most bibliometricians are accustomed to navigating these errors in their analyses, and some even know how to fix them upstream (e.g., curating ROR records and string matching in OpenAlex). On the other hand, research security professionals come from diverse backgrounds and only sometimes have a nuanced technical understanding of bibliometric databases. That means an affiliation error that used to be a minor data-quality annoyance can now hold up a grant, trigger a foreign-influence review, or put a researcher into a compliance inquiry over a relationship that never existed.

A first pass at minimizing false positives

We took the published watch lists that screening actually runs against (Canada’s Named Research Organizations, the US Department of Defense 1260H list, the US Consolidated Screening List including the BIS Entity List, US OFAC, the US UFLPA list, the UK FCDO sanctions list, and Australia’s DFAT list) and cross-referenced each entity to ROR with a strict country-consistency check. That produced a verified crosswalk of 192 organizations covering roughly 4.24 million works.

Then, for each organization, we screened every affiliation string linked to it, about 2.5 million in all, looking for this specific failure: an unrelated company or institution absorbed into a listed entity because their names collide (often with an organization lacking a ROR record). We found 6,191 such over-merges, roughly a quarter of one percent of those strings, and we submitted manual curations to remove the linkages. The scale of the collision problem is easier to see in a few real examples than to describe.

• A small US biotechnology firm, identified on its papers by a short name only that also refers to a broad class of medical interventions, was linked to a Russian research institute whose registered acronym is nearly identical. The two have nothing to do with each other. The firm simply had no ROR record, so the closest-looking acronym won.
• System Equipments Div., Asahi Engineering Co., Ltd, a Japanese engineering company, was linked to a Chinese entity on the BIS Entity List whose registered name is essentially “System Equipment.”
• ENGINEERING-PHYSICS, INC. was linked to a Russian Institute of Engineering Physics that appears on these lists, on the strength of the shared words alone.

None of these organizations has any connection to the listed entity it was linked to. In each case the two names are close enough that you can see why an algorithm reached for the wrong one because no record existed for the real organization. This work helps reduce these errors, but they still exist.

Advice to analysts doing the screening

If you’re an analyst in charge of screening research applications for potential connections with a list of organizations of concern, never assume that a bibliometric match is a definitive signal of a connection. A match can help you shorten your list of connections to confirm manually, but there are errors and you need to confirm any potential connections. You can do that easily by inspecting the raw_affiliation_string field in the OpenAlex API. For instance: https://api.openalex.org/works/w3134510327 -> raw_affiliation_string : “Optical Sciences, University of Arizona, Tucson, Arizona, USA”. You can also visit the publication’s landing page or PDF for more information directly from the source. All of that information is linked conveniently in the OpenAlex record.

You should also take efforts to identify false negatives. For instance, if the organization of concern does not have a ROR record (or doesn’t have enough information on how the organization represents itself in publications), then we won’t be able to make any matches to that institution. More broadly, I highly recommend that you connect with experts in this field on how to successfully leverage bibliometric tools in your work while respecting their limitations.

Advice to researchers submitting to screenings

Take a look in OpenAlex in advance to see if there might be issues in your author profile (my profile below). Sometimes another authors’ works might have snuck into your author profile which you can edit that there directly. And sometimes there are errors in affiliation attribution even when the work is yours (like in the examples above). When that happens, the quickest way to get it sorted out is to send details in an email to support@openalex.org. But keep in mind that funders and analysts use different databases and tools for research security compliance workflows and many of them use closed, proprietary databases you won’t be able to access.

Let’s work on this together

If you work in research security, or at a funder or institution that screens affiliations, we want to hear from you.

Screening decisions carry real weight, so the data behind them should be open to inspection. In OpenAlex every affiliation link traces back to the source publication, every correction is public and permanent, and the code that makes the matches is on GitHub (github.com/ourresearch). Nothing sits in a proprietary index that no one outside the vendor can check. When a match is wrong you can see why, contest it, and get it fixed, and that fix then helps everyone who screens after you. For a decision that can affect a researcher’s funding or reputation, that is not a small difference. Open access to the data also means scrutiny is not gated by budget: a well-resourced agency and an under-resourced one work from the same public record.

We do not think we have this solved. We know some of the limits and we are certain there are more we have not found. What we want is to work with people who do this screening every day to map those limits honestly, so nobody treats a match as more certain than it is, and to build the documentation and tooling that would make an open, auditable corpus genuinely good infrastructure for this work. A few experts collaborating with us in the open could improve it for the whole community, globally.

Get in touch at kyle@openalex.org.

TL;DR. OpenAlex is getting steadily better at linking research publications to the funders and grants behind them. But the methods that make that possible aren’t working as well for datasets. And we’re hearing from funders around the world that understanding the datasets they’re funding is critical to them. That gap is where DataCite’s funding metadata becomes essential, and OpenAlex now ingests it directly: funder and grant links on 716,297 dataset records. Put differently, DataCite supplies 96% of every dataset-funder link OpenAlex holds. DataCite’s work connecting funders and grants to research outputs is quietly becoming critical infrastructure for open funding intelligence.

We’re getting better at finding funders. Except for data.

We’re less than a year into our Wellcome-funded project to integrate research funding metadata into OpenAlex and we’re getting good at linking outputs to the underlying funders and grants. For journal articles, the share carrying both a funder and a specific grant number has climbed from about 6% of 2015 articles to roughly 15% of 2025 articles, which in absolute terms is more than 10 million articles published since 2015 linked to the grants that paid for them. Preprints and reviews are on the same upward path.

Then you get to datasets.

Figure 1. Share of each output type whose record carries both a funder and a grant number. Datasets, and software, sit near zero.

Just 1% of datasets published since 2015 carry a funder and grant, and unlike the article curve, that number is not climbing. These are not niche outputs. They are exactly the research products funders are asking us to help them track, and they are the ones where our coverage is thinnest.

A dataset has nowhere to say thank you

The reason is structural, and it is worth being precise about.

Most of what OpenAlex knows about who funded a piece of research, we learn from the paper itself (or directly from funders). Articles have an acknowledgements section, and it usually names the funder and the grant. We read those sections at scale, and it works: full-text mining has allowed us to assert millions of linkages between outputs and funders. 

Datasets do not have acknowledgements sections. There is no paragraph at the end of a Zenodo deposit thanking the European Research Council. So the single most productive technique we have for articles produces almost nothing for data. It is not a tuning problem or a coverage problem. The text simply is not there to read. Researchers also seem to report datasets less frequently than research articles in their reports to funders (which we can also ingest). 

That leaves exactly one place where the funding of a dataset can be recorded: the structured metadata supplied when the dataset is registered and gets its DOI. And that is precisely what DataCite built. Their metadata schema treats funding as a first-class field, fundingReferences, with room for the funder’s name, a persistent identifier for the funder, the award number, and the award title. It is a deliberate, unglamorous piece of infrastructure work, and it is the only reason the question in this post’s title has an answer at all.

Where dataset funding data actually comes from

We looked at every funder link on a dataset in OpenAlex and asked a simple question: which source told us about it?

Figure 2. Of the 716,297 funder-linked datasets, the share whose link each source supplied.

Of the 716,297 datasets in OpenAlex that carry a funder link*, DataCite asserted 685,778 of them, about 96%. Crossref accounts for around 12,000, and our own methods, the full-text mining and the funder grant records we ingest directly, for around 10,000 more. 

This is not a story about one registration agency outperforming another. Crossref and DataCite serve different corners of the scholarly record, and each does the job its corner needs. The point is narrower and more interesting: for research outputs that are not articles, structured funding metadata at registration may be the only viable path forward. 

One grant number, and everything it unlocks

Here is what makes this worth the effort.

The Mammal Diversity Database is deposited on Zenodo. Its DataCite registration declares one funder and one award number: the US National Science Foundation, grant 1441737. That is all the dataset itself says.

But OpenAlex already holds NSF grant 1441737, ingested directly from NSF’s own award data. So the two fuse, and the dataset inherits the entire grant record:

Collaborative Research: VertLife Terrestrial: A complete, global assembly of phylogenetic, trait, spatial and environment characteristics for a model clade

Principal investigator: Walter Jetz  ·  $1,130,504  ·  1 October 2014 to 30 September 2020

None of that was in the Zenodo record. All of it is now attached to the dataset, because a depositor typed a grant number into a form. This happens at scale: 190,361 datasets, about 28% of the ones in DataCite with funder information, connect to a full grant record from the funder’s own data this way– and that number will only grow as OpenAlex continues to ingest grant records directly from more funders. 

The links also handle real-world complexity. A dataset on nuclear transport deposited in 4TU.ResearchData carries three grants from two funders: two from the Dutch Research Council, including “BaSyC – Building a Synthetic Cell,” and a European Research Council grant. The examples span every field and region we looked at, from an Australian Research Council grant behind a linguistics dataset on word order in Malayo-Polynesian languages, to a VINNOVA grant behind a Swedish mammography dataset built for AI research, to an ANR grant behind seismic data from a Réunion hotspot experiment.

The field is there but it is currently mostly empty.

Now the part we should all be less happy about.

The reason dataset coverage is only at 1% is not mainly that OpenAlex is missing links that exist. It is that the links were never registered in the first place. Even among the curated repositories where a researcher fills in a deposit form, the funder field usually stays blank: Zenodo sits around 8%, Figshare around 2%, Harvard Dataverse near zero.

The one clear outlier is the US Department of Energy’s OSTI, at 57%. Nearly all of that comes from national laboratories whose deposit workflows already know which DOE program paid for the work, so the field gets populated automatically. That is the whole lesson in one data point: when the deposit workflow captures the funder, the funder gets recorded. Where it is an optional free-text box at the bottom of a long form, it stays empty.

We can also see that when depositors do fill it in, they do it properly. Of the funder-linked datasets, 696,626, or 97%, carry a specific grant number and not just a funder name. People are not being vague. They are just not being asked.

What we built

As part of a project funded by Wellcome to build out open funding metadata, OpenAlex now ingests DataCite’s funding references directly. It went live on 18 June 2026 and refreshes every day, so new registrations flow through without waiting on a quarterly rebuild.

You can query it now. Every funder-linked dataset:

https://api.openalex.org/works?filter=type:dataset,funders.id:!null&include-xpac=true

Or everything a given funder has data-wise. Wellcome, for instance, has 2,432 datasets linked in OpenAlex today, and the NIH has 10,538:

https://api.openalex.org/works?filter=type:dataset,funders.id:F4320311904&include-xpac=true

All of it is in the API, the snapshot, and the web interface, under the same CC0 terms as everything else in OpenAlex.

What this means for you

If you run a repository, or publish datasets: add funder and award number to your deposit workflow and pass them to DataCite as fundingReferences. Even just the funder name and award number as text strings is enough. This is the single highest-leverage metadata change available to you right now, and the OSTI numbers show what happens when you make it.

If you are a funder with a data management policy: add a line to it. Ask your grantees to name the funder and the award number when they register a dataset with DataCite. That one sentence in your policy will do more for your ability to see and evidence your own data outputs than any amount of downstream text mining that we or anyone else can do, because for datasets that text mining structurally cannot work.

If you are a researcher depositing data: it is about thirty seconds at deposit time, and it is the difference between your dataset being connected to the work that paid for it and being an orphan.

Finally, our thanks to DataCite. Making funding a first-class field in the metadata schema, and doing the long, patient work of connecting funders and grants to research outputs, is exactly the kind of open infrastructure that is invisible when it works and impossible to substitute for when it is missing. As funders, institutions, and the public increasingly ask not just what research was published but what it produced and who paid for it, that connective tissue becomes the foundation of open funding intelligence. Everything in this post rests on it.

Spot something that looks wrong, or want to talk about connecting your funding data to OpenAlex? We would love to hear from you: support@openalex.org.

—

* These figures leave out a small number of very large automated registrations, where one funder name is repeated across millions of machine-generated records. Because those reflect bulk data series rather than individual funding decisions and dwarf all other signals, we have set them aside for now, and will revisit how best to represent them.

This week we replaced how OpenAlex decides a work’s type. Every work we ingest carries this field — article, book-chapter, dissertation, dataset, retraction, and so on (the full taxonomy is here). A new classifier now assigns every work’s type from the evidence in its own record, using an auditable set of rules, and on its first production run it changed the type of 49.6 million works — about 10% of the catalog.

(I’m Rohan Mantena, a new data/software engineer on the team, and this was our team’s project and my first big one here. Excited to share it.)

Why work types matter

Correct type classification matters to our users for a lot of reasons. A few we want to highlight:

• Finding the right research outputs: the actual papers and reviews for a literature search, the datasets to build new research on, the preprints for tracking early results.
• Excluding what isn’t relevant: most searches don’t want tables of contents, editorial boards, or the photographs and physical objects that repositories legitimately hold (paratext and other exist so you can filter them out).
• Citation metrics: Field-Weighted Citation Impact (FWCI) normalizes each work’s citations against works of the same type, field, and year. When conference papers and editorials are mislabeled as articles, they land in the wrong comparison group.

This is an inherently difficult problem. A work’s type was largely inherited from its source at ingestion, and sources genuinely differ: Crossref, DataCite, and institutional repositories each maintain their own type classification scheme, so the same kind of work arrives under different labels depending on where it came from. And even setting the schemes aside, there’s no verification of what a depositor selects — a type label is whatever someone chose from a dropdown, years or decades ago. Many sources provide no type at all. Different vocabularies plus unverified labels meant article became the default dumping ground, and our users noticed: support tickets flagging classification errors are part of what motivated this work.

So, we set out to fix it.

Execution

As with any good classifier, we needed good data and a ground truth.

A taxonomy. Before labeling anything, we had to agree on what the labels mean: is a book review a ” review”? Is a journal issue’s front matter a work? We settled on 25 canonical types (full taxonomy and definitions at openalex.org/types). Casey wrote annotator-grade definitions for each, with inclusion examples, exclusions, and the distinguishing rule against each type’s nearest neighbor. Kyle and Jason settled the contested boundary cases, like where paratext ends and other begins. Every classification decision downstream traces back to those definitions, and boundary questions get a ruling there before they get code.

A gold set. We sampled works from across the catalog and built an LLM-labeled gold set: 52,383 works, plus a 1,495-work validation set. Because LLM labels deserve scrutiny, we audited them: models from different model families independently labeled the same works, and their agreement rate bounds the labeling error — which turned out to be low. The gold set also gave us a rough measure of the problem: existing types matched it only about 70% of the time.

Building the ruleset. With solid footing, the error and rule patterns became much clearer. Repositories announce dissertations in dozens of languages (Tesis, Hochschulschrift, väitöskirja), and each vetted spelling became a rule. Correction notices start with “Erratum to:”. Conference abstracts hide in journal supplement issues. Kyle personally adjudicated over 600 disputed works, and his rulings became rules as well. Each rule states its evidence, its target type, and its guard conditions.

The effort was worth it. The output is an ordered cascade of ~160 rules, compiled into a single SQL expression that re-runs over the full corpus every day. Rules are ordered by measured trustworthiness: unambiguous structural signals fire first, softer text heuristics later, and the first match wins. If nothing speaks, the work is typed article — that default has a cost, and we come back to it below.

Why rule-based? Openness and improvability. Every classification is explainable — directly attributable to a specific rule — and can subsequently be improved. And the taxonomy stays governable: when we make a ruling about what counts as paratext vs other, that ruling maps onto specific, inspectable rules.

How we measured it

Checking the classifier against the existing types would have been the wrong test: it would simply have measured conformity to labels we already suspected. Early in the project, one improvement made our agreement score go down, and inspection showed the classifier was right and the old labels were wrong. So we evaluated against independent ground truth instead, two ways.

A blinded truth panel. We sampled 1,000 works and had three LLMs from different model families label each one. They were shown the work’s evidence but blinded to both the current OpenAlex type and the classifier’s prediction, so there was nothing to anchor on. The labelers agreed with each other strongly (κ = 0.90), and disagreements were resolved by majority. Result: the old types were right 69.3% of the time, the classifier 78.1%. And on the works whose type actually changed, the panel found 43 flips from wrong to right and 0 from right to wrong — where the classifier disagrees with the old data, it won every audited case.

An admission bar for every rule. Aggregate accuracy can hide a bad rule inside a good system, so before admission, each rule is tested on a blinded sample of the works it would actually fire on in production, against a ≥95% precision bar. Rules that miss the bar get fixed, guarded, or shelved with their evidence attached.

The results

Here is what the first production run did to the catalog:

type	before	after	change
conference-paper	2.8M	18.2M	6.4×
conference-abstract	535k	4.2M	7.8×
book-review	80k	2.1M	26×
editorial	833k	1.3M	1.6×
review	4.2M	1.0M	-76%
erratum	420k	625k	+49%
retraction*	23,142	45,454	2×
preprint	10.1M	9.3M	-8%
article	264.5M	240.3M	−9.1%

* Retraction as a type refers to our ability to detect when a record is a notice of retraction, which is separate from when an existing work is retracted

Changes to the type list

Three vocabulary changes came along, relevant to anyone filtering by type:

• software is new —  (~1M works) stay tuned for more on this soon 😉
• software-paper is new — a small type (~9,300 works) for papers describing software, distinct from software itself.
• report-component is gone — its ~800 works are now simply report.
• book-section is gone — its 2,000 works were mostly index/front-matter records, now reference-entry or paratext.

If your pipeline filters on either retired value, it will now return zero results; please update to the successor types.

What’s still hard

Our classifier isn’t perfect:

Types that leave no trace. A letter or editorial that the publisher registered as a plain journal article, with no distinguishing title, is invisible to a rules engine and stays article. Precision on these types is high; recall is the headroom. Those works are labeled no worse than before, and the fixes are input problems more than rule problems: we’ve parsed the metadata embedded in roughly 300 million landing pages, and more evidence channels are on the way.

Works with no usable metadata default to article. About a quarter of works carry no raw type at all. When nothing is indicative, article is our stated default, and when the previous label was more specific, defaulting can be a regression. We measured this tradeoff before accepting it, and chose a transparent, reproducible default over preserving labels of unknowable provenance. We monitor exactly the flows where it is induced, and recovering those works from actual evidence is on the roadmap.

Coverage varies by source. Crossref-registered works carry the richest signals; repository and legacy records are poorer, and their types will improve as more of their vocabulary passes the admission bar.

All that said, the classifier is built to close its own gaps. Every limit above is a future rule or evidence source, each one is tested against the admission bar before it ships, and because types are recomputed from current evidence daily, every fix reaches the whole catalog the next day. As we adapt and add rules, the numbers in this post keep getting better. And this approach isn’t staying confined to types: we plan to roll out the same rule-based, auditable pattern across other metadata fields over the next year — source is next in our sights.

What this means for you

• Roughly 10% of works changed type in the snapshot and API this week; the daily changefiles carry the updates. If you cache types, expect a large one-time diff.
• Counts of conference papers, errata, retractions, editorials, and book reviews are substantially higher than in any previous OpenAlex snapshot.
• If you compute or consume FWCI-style indicators, expect small shifts as tens of millions of works move into their correct comparison groups.
• If a type looks wrong, submit a ticket to support@openalex.org. Every verdict is attributed to a specific rule, so a good report (“this work is typed X, should be Y”) points straight at a rule we can fix — and fixes propagate to the whole corpus the next day. Systematic patterns are even more valuable: a single repository vocabulary we’re misreading can be thousands of works.

Your error reports made these improvements possible, so thanks again for sending us issues when you see that. Please keep ’em coming! We’re looking forward to continuing to work with the community the increase the quality and accuracy of open research information!