When a paper thanks "the National Institutes of Health," current tools record one fact: which organization paid. They do not record what kind of support it was. The money could have been a single-project grant, a shared-equipment award, access to a core facility, a multi-site clinical network, or a student fellowship. That distinction matters for anyone studying how unstable funding shapes science. Studies of long-running research networks describe the same pattern again and again. What breaks first when budgets wobble is often not the scientific idea but the supporting infrastructure it depends on. The Pediatric Emergency Care Applied Research Network reports operational compromises under extramural funding instability (Chamberlain et al.), and the PAGES paleoscience network is described as surviving organizational and funding shifts (Zhao and Yan). The literature on regional development makes the same argument about roads and power grids: a valuable primary asset underperforms when its auxiliary infrastructure is missing (Cheng; Sugandi et al.). The acknowledgment section is the one place where science routinely writes down its own support structure, yet it is read only at the coarsest level.
We propose to close this semantic granularity gap, meaning the mismatch between broad category labels and the specific labels that downstream reasoning needs, for research funding. We will build a small hierarchical ontology of funding instruments. We will train an extractor that assigns each acknowledged support mention to a node in that ontology, supervising it cheaply through a funder whose grant numbers already encode the instrument type. Finally, we will test whether instrument-level records predict the continuation or truncation of fine-grained research topics better than organization-level records do. The contribution is not another entity-typing model. It is a measurement instrument, validated by its ability to reveal an effect that coarse funding records hide.
Core Claim
Our claim has two parts, and each can be tested separately. The first is that the type of a funding instrument can be recovered from the text of an acknowledgment for funders whose grant identifiers carry no type information. This holds if a model trained only with labels derived from identifier codes transfers to hand-annotated acknowledgments from other funders, with the identifiers masked. The second is that instrument type carries information about research continuity that funder identity does not. Specifically, we hypothesize that the disappearance of infrastructure-type support (equipment, facilities, networks, compute) is followed by truncation of methods-heavy subtopics more often than the disappearance of project-type support. Topics are measured with the Computer Science Ontology's roughly 26,000 topics (Salatino et al.) and a domain-appropriate equivalent where needed. If the first part fails, the second cannot be run at scale. If the second fails, instrument typing is a curiosity rather than a tool.
Research Program
The closest prior work is Find the Funding. It detects funder mentions in acknowledgments with a BERT model given extra pretraining on acknowledgment text. It then links those mentions to Crossref funder records or to a "not in the knowledge base" label, using a two-tower encoder called FunD plus a gradient-boosted reranker. It also releases the ELFund and EDFund benchmarks. Its output is an organization identifier and nothing more. We reuse its mention detector and linker as the first stage and add a second stage that FunD does not attempt: typing each support mention by instrument, and typing each paper by fine-grained topic.
The ontology will be deliberately small, on the order of a few dozen nodes in a strict parent–child hierarchy. The branches will be project support (investigator-initiated, cooperative agreement), people support (fellowship, training grant, career award), infrastructure support (equipment, core facility, compute or data allocation, consortium or network), institutional support (center or block funding), and in-kind support. We keep the hierarchy strict because OntoType found that restructuring the FIGER ontology into strict hypernym–hyponym form alone added 2.1 macro-F1 in zero-shot typing. That result constrains our design toward a clean tree rather than an open label vocabulary. It also matches the finding in Ultra-Fine Entity Typing that open vocabularies produce contradictory assignments across branches.
Supervision follows the distant-supervision recipe of FIGER, which labeled two million sentences by linking Wikipedia anchors to Freebase types. We do the analogous thing with grant identifiers whose format encodes the mechanism. NIH activity codes are the obvious source, since their prefixes distinguish research projects, fellowships, training, career, center, cooperative, and shared-instrumentation awards. These codes give us labels for free wherever an NIH number appears. FIGER also shows the cost of this recipe: labels inherit the source's errors. We will therefore measure label noise on a hand-checked sample.
The model will be a fine-tuned encoder that classifies over the hierarchy. It will be trained with the identifier string masked, so that it must learn from surrounding language ("instrument purchased under," "access to the core," "network sites"). For transfer we add an ontology-guided retrieval variant in the style of ORAG, which retrieves structured descriptions of parent and child ontology nodes rather than raw text chunks.
For the downstream analysis we assemble a longitudinal corpus of papers with acknowledgments, starting from the pre-2017 ELFund papers and extending with openly licensed full texts. Each paper is typed by topic with the CSO classifier. We group papers into author-cluster trajectories and define an instrument-lapse event: an instrument acknowledged across a cluster's papers stops appearing. We then ask whether that cluster's specific subtopics continue afterward. This is observational. We will not claim causation. The design goal is a comparison between coarse and fine representations under identical confounds.
Baselines
The naive reference point is organization-level tracking alone: FunD-style funder identifiers with no instrument typing. For the downstream question, this is the "do nothing new" system that bibliometric studies already have. For the extraction question, the strongest existing linker is FunD from Find the Funding. Its authors report best F1 on EDFund in both the in-knowledge-base and emerging-entity splits and roughly four-fold speed over the boosted baseline. However, the file gives no numerical F1, so no reported result was found on Lacuna for FunD's ELFund or EDFund scores. We will rerun it from the released setup and report our own number before building on it. Beating it is not our goal. Matching it on funder linking confirms our first stage is sound.
For instrument typing itself, no prior system exists, so the strongest competitors are general typing methods applied to our ontology. The first is ORAG, which the file reports lifts F1 by 10% to 24% over vanilla retrieval-augmented prompting on the BattINFO and Genshin Impact typing benchmarks. That is the best reported evidence for ontology-guided retrieval with language models on specialized, narrow ontologies, and it is the most likely method to transfer. We will run it with the same backbone language model, the same ontology descriptions, and the same retrieval budget as our variant. The second is OntoType, which requires no annotation and reports +14.6 macro-F1 over ZOE on OntoNotes. We will run it zero-shot on our ontology.
The deterministic rule baseline parses grant identifiers and maps known codes to instrument types. It will be scored only on mentions that carry parseable identifiers, and separately on all mentions, where it abstains elsewhere.
The key ablation removes the text context and keeps only the identifier. A second ablation removes the identifier and keeps only the text. Together they isolate whether our model learned transferable language cues. Beating the baselines means higher macro-F1 than ORAG on the non-NIH held-out set under equal tuning budgets, measured on development data drawn from the same funders as the respective training source. For the downstream analysis, it means instrument features adding predictive value for subtopic continuation beyond funder features.
Why not just…?
Why not just parse grant numbers, since many already encode the mechanism? This is the most serious threat, and it is why identifier codes are our supervision source rather than our competitor. For NIH-funded papers, a regular expression may do nearly as well as any model, and we will report that honestly. The rule has two gaps. It is silent on funders whose identifiers do not encode type. It is also silent on the many infrastructure acknowledgments that carry no number at all ("we thank the imaging facility"). The decisive test is the rule's coverage on the hand-annotated non-NIH set against the masked-identifier model's F1 on the same set. If coverage is high and the model adds little, the extraction contribution collapses to a lookup table. We would then report the downstream analysis restricted to coded funders.
Why not just prompt a large language model with the list of instrument types? Evidence from ORAG suggests vanilla retrieval-augmented prompting hallucinates on specialized ontologies, and ontology-structured retrieval gains 10–24% F1. That evidence comes from battery materials and a video game, not from acknowledgments, whose phrasing is formulaic and may be easy for a language model. We therefore include plain prompting as a baseline. If it matches the fine-tuned model, we will adopt it as the extractor, because the downstream claim does not depend on which extractor wins.
Why not just use coarse field codes instead of 26,000 CSO topics for the downstream analysis? Instrument lapses plausibly truncate a specific method, such as one imaging modality, while the broader field keeps publishing. Coarse codes would average that away, just as coarse typing buries half of OntoNotes mentions in "Other" (Choi et al.). This is itself a testable prediction. We will run the continuation analysis at three topic depths (field, CSO mid-level, CSO leaf) and expect the instrument-type effect to grow with depth. If the effect appears equally at the field level, the fine-grained topic layer is unnecessary.
Proposed Experiments
The first experiment tests extraction transfer. We train on masked-identifier NIH acknowledgments and evaluate on a hand-annotated set of several hundred acknowledgments from non-NIH funders, drawn partly from ELFund papers. We report macro- and micro-F1 at each ontology level, the rate of hierarchy violations, and agreement with a second annotator. We also report the identifier-only and text-only ablations. Success means the text-only model beats the rule baseline's full-set score and is competitive with ORAG.
The second experiment is the downstream decisive check. We predict whether an author cluster's leaf-level CSO subtopic persists two or more years after an acknowledged instrument lapses. We compare three logistic models: funder-only features, funder plus instrument type, and funder plus a shuffled instrument-type control. Each is fit under identical controls for cluster size, publication rate, and topic age, and evaluated by held-out log-likelihood and area under the curve. The hypothesis survives only if instrument type improves held-out prediction over both the funder-only and shuffled models. It also requires that the infrastructure branch show a larger truncation association than the project branch, and that this difference strengthen from field-level to leaf-level topics.
A negative result at the leaf level with a positive one at the field level would falsify the granularity argument. A null everywhere would indicate that acknowledgments, as written, do not record the instability effects documented qualitatively by PECARN and LTER researchers. We would report that as a finding about the limits of acknowledgment-based science-of-science.
References
- The Pediatric Emergency Care Applied Research Network: Progress and Update
- Bridging global paleoscientists: an interview with PAGES Executive Director Marie-France Loutre
- Research on the sustainable development policy of rural tourism in Ganzi Prefecture
- wisata Pengembangan Desa Wisata di Kelurahan Kemumu Kecamatan Arma Jaya Kabupaten Bengkulu Utara
- Classifying Research Papers with the Computer Science Ontology
- Find the Funding: Entity Linking with Incomplete Funding Knowledge Bases
- OntoType: Ontology-Guided and Pre-Trained Language Model Assisted Fine-Grained Entity Typing
- Ultra-Fine Entity Typing
- Fine-Grained Entity Recognition
- ORAG: Ontology-Guided Retrieval-Augmented Generation for Theme-Specific Entity Typing
- Graduate students navigating social-ecological research: insights from the Long-Term Ecological Research Network