The bacteria at the center of the antibiotic resistance crisis do not respect academic boundaries. Resistance genes shuttled by plasmids and transposons move freely between hospital wards and rivers — Philipp Köhler et al. found near-identical metallo-β-lactamase genes in patients and local aquatic environments. The literature about those bacteria, however, is organized by discipline: clinical infectious disease, environmental microbiology, veterinary and agricultural surveillance, and computational genomics each publish in their own journals, in their own vocabularies, and largely cite themselves. The study of how such disciplinary boundaries form, blur, and fragment — Wu's analysis of data mining's contested multi-parent origins and Neumann et al.'s work on vocabulary fragmentation in young fields — has never been turned on antimicrobial resistance (AMR), even though AMR is arguably the field where the cost of that fragmentation is measured in deaths.
This matters concretely because two documented bottlenecks in the AMR response could be symptoms of exactly this siloing. First, newly characterized resistance genes stay trapped in unstructured papers instead of entering reference databases like CARD, a curation bottleneck that SCIARG tries to automate. Second, antibiotic developers keep rediscovering known compounds, wasting part of the $>1$ billion and $>10$ years each drug costs, which an LLM-and-knowledge-graph alert pipeline attacks with entity linking. Neither work asks why the relevant knowledge is missed — but Wahle et al. have shown that fields forget their older literature ("citation amnesia," a 12.8% collapse in citation age in NLP) and cite external disciplines very differently than internal ones. If AMR sub-communities read past each other, the curation lag and the rediscovery problem become measurable consequences of disciplinary boundary ambiguity, not just engineering inconveniences.
Core Claim
The claim is that citation and vocabulary silos across AMR sub-communities are a measurable driver of curation lag and of automated-curation errors, and that making literature mining boundary-aware — conditioning it on the publishing sub-field and normalizing its terminology — closes a specific, diagnosable part of the error. The proposal has two halves that test each other: a bibliometric diagnosis establishing that the silos exist and predict outcomes, and an intervention on the SCIARG benchmark showing the diagnosis points at a fixable failure mode. The silos are quantified with a simple internal-citation fraction, where $C_{i,j}(t)$ counts citations in year $t$ from papers in sub-community $i$ to papers in sub-community $j$:
$$S_i(t) = \frac{C_{i,i}(t)}{C_{i,i}(t) + \sum_{j \neq i} C_{i,j}(t)}$$
so $S_i = 1$ means a community cites only itself. The hypothesis is that $S_i$ has risen over 1980–2023 even as the physical exchange of resistance elements between those communities' study systems intensified, and that resistance assertions published far from the curation community's home field are curted latest and extracted worst.
Research Program
The first task is building a subfield-resolved corpus of the AMR literature. Notten et al. showed that keyword-based corpus construction in hybrid fields produces high false-positive and false-negative rates, because fields defined by subject matter rather than method have no natural bibliometric perimeter. So instead of searching keywords, we grow the corpus by seed-citation expansion: start from the source papers behind the SCIARG statement corpus and known venue anchors for each sub-community, expand one co-citation hop at a time, and cluster into sub-communities (clinical, environmental, veterinary, computational genomics) using venue and citation structure. This runs on OpenAlex, the same ~240-million-paper, 280-billion-citation repository that Wahle et al. used for their 20-field, 43-year citation analysis; we reuse their mean Age of Citation and Volume-Adjusted Average Citation Age metrics at sub-community granularity, which their aggregate "Medicine" and "Biology" categories cannot resolve.
The second task is diagnosis. We compute $S_i(t)$ and per-community citation ages over 1980–2023, testing whether AMR shows the asymmetric pattern Wahle et al. found at the field level — recent internal citations, older external ones — and whether corporate and educational citation horizons diverge here too. We then apply the Lakatosian mapping of Kuskova et al., who used fractional word co-occurrence networks and their "island" sub-graph method on the journal Network Science to show that an interdisciplinary field can form an immutable conceptual "hard core" (49 core nodes) buffered by protective belts of applications. Applied to AMR, this answers a question their network-science study could not: does the "One Health" framing that Köhler et al.'s environmental-clinical gene flow implies actually exist as a shared conceptual core, or are environmental, veterinary, and clinical AMR research separate islands sharing only a slogan? Our prediction, grounded in Neumann et al.'s finding that multi-parent fields develop fragmented vocabularies, is the latter.
The third task links the bibliometrics to consequences. For each gene–resistance assertion in the SCIARG gold standard we measure curation lag — years between the first supporting publication and the gene's entry into CARD — and regress lag on the publishing sub-community's citation distance from the curation community and on vocabulary overlap with curated nomenclature. No reported result found on Lacuna for curation lag in reference databases, so this is a first measurement, not an improvement on one.
The fourth task is the intervention. We retrain the SCIARG statement-extraction model — a PubMedBERT model deciding whether a sentence asserts that a gene confers resistance — with two boundary-aware ingredients. First, sub-field conditioning: the model is told (or infers) which community's writing conventions a sentence comes from, so "confers resistance" in a clinical trial report and an ectopic-expression molecular paper are scored against community-specific framings. Second, a cross-community synonym map built from the co-occurrence networks of task two, so that environmental and veterinary terms for the same determinants map to shared representations. The contrastive idea is borrowed from Lu et al., who showed that when candidate categories crowd together and semantic boundaries blur, explicitly contrasting confusable categories (their Reduce-then-Compare with Pairwise Contrastive Chain-of-Thought lifted zero-shot GPT-3.5 from 40.24% to 62.01% accuracy on fine-grained classification) resolves fuzzy perimeters; here the confusable "categories" are sub-field framings of the same assertion.
Proposed Experiments
Experiment 1 tests the silo hypothesis on the OpenAlex corpus: per-community citation ages, $S_i(t)$ trajectories, and the island-structure analysis. Decisive checks are whether $S_i(t)$ rises after volume adjustment (Wahle et al.'s VACA controls for publication growth, and they found recency bias survives it, with only a 0.19 correlation between volume and citation age in NLP), and whether AMR sub-communities' citation ages fall at rates comparable to the −3.4% cross-field average or worse.
Experiment 2 tests the curation-lag consequence: lag should increase with sub-field citation distance and decrease with vocabulary overlap. If lag is uniform across sub-fields, the boundary story is wrong and the proposal should be abandoned at this stage — that is a genuine risk, and this is where it is falsified cheaply.
Experiment 3 is the intervention, evaluated on the SCIARG gold standard, stratified by source sub-field (characterizing that distribution is itself a deliverable). Primary metric: F-score and recall overall and per stratum. Secondary metric: the F-score spread between best- and worst-represented sub-fields, which the original paper never reports and which we hypothesize is large.
Experiment 4 tests the rediscovery link: we re-run the LLM–knowledge-graph alert pipeline on its 73-organism candidate list and ask whether the organism–chemical–activity evidence triples it retrieves late or with weak alerts are disproportionately old and cross-disciplinary in citation terms. The pipeline retrieved 26 of 27 known triples, so coverage of known evidence is nearly perfect; the question is whether what it misses or flags weakly carries a citation-age and sub-field-distance signature.
Baselines
The naive reference point is the heuristic keyword and rule baseline of SCIARG: dictionary matching for phrases like "conferred resistance," which scored an F-score of 47.4% with only 32.2% recall on the gold standard, failing both because resistance assertions are linguistically diverse and because keywords cannot distinguish phenotypic resistance from contextual qualifiers like ectopic expression. This defines the floor; any boundary-aware method must obviously clear it, and it usefully embodies the "just use keywords" position that also underlies keyword-based corpus construction, which Experiment 1 replaces with citation expansion.
The strongest competitor is the SCIARG ensemble of PubMedBERT models with prompt tuning and masked-language-model pre-training, which achieved an F-score of 77.0% and 89.0% recall on the gold standard. We will re-run this ensemble ourselves under identical conditions — same gold and silver data, same PubMedBERT backbone, same prompt-tuning and MLM budget, same compute — so that our only change is the boundary-aware conditioning and synonym mapping. Beating it means two things jointly: matching or exceeding 77.0% overall F-score, and cutting the worst-sub-field to best-sub-field F-score spread by at least half. An overall F gain with an unchanged spread would indicate the fix is generic capacity, not boundary awareness, and would count against the claim. For the bibliometric half, the baseline expectation is the broad-field behavior documented by Wahle et al.: if AMR sub-communities behave like the aggregate biomedicine their metrics imply, there is nothing AMR-specific to explain.
The key ablation isolates the proposal's central ingredient: the same conditioned architecture with the sub-field signal replaced by a random or constant label, and the synonym map replaced by identity. If the full model beats this ablation on out-of-community statements specifically, the boundary mechanism — not the extra parameters or extra training signal — is doing the work. This matters because Kuskova et al. found that methodological complexity has an inverted-U relationship with impact (β ≈ −0.40 on the squared complexity term in their citation regressions), so the parsimonious mechanism should win the argument, not the heavier one.
Why not just…?
Why not just scale the curation model — a bigger LLM with retrieval augmentation? This is the real threat to the intervention, and we do not argue around it. A frontier model with retrieval might lift extraction accuracy for reasons unrelated to disciplinary boundaries. The stratified error analysis of Experiment 3 is designed to settle it: if the SCIARG ensemble's errors are uniformly distributed across sub-fields and shrink uniformly with scale, the failure is capacity, the boundary framing is wrong, and we will say so. But if errors concentrate on statements from under-represented sub-fields and involve terminology mismatch, scale is the wrong lever — and GenoARM's result that a 5-gene panel plus metadata reaches 93.77% accuracy on E. coli resistance, matching panels of ~200 genes, is a precedent in this very literature for parsimonious, targeted signal beating brute-force coverage.
Why not just extend the LLM–knowledge-graph alert system to more databases? More entity linking and more databases would improve retrieval coverage, but the pipeline already retrieves 26 of 27 known evidence triples — coverage of what is known is close to saturated. The open question is why the underlying knowledge was invisible to researchers for years, which is a claim about citation behavior, not database joins. Only Experiment 4's design — regressing detection latency on citation age and sub-field distance — can test it, and no reported result found on Lacuna for any study linking citation behavior to the antibiotic rediscovery problem.
Why not just publish the bibliometric study and skip the machine learning? A measurement alone invites the correlation-is-not-causation objection: maybe siloed communities and slow curation are both effects of something else. The intervention is what converts the diagnosis into evidence of mechanism — if conditioning on the very variable the regression identifies (sub-field framing) repairs the errors the regression predicts, the causal story is far stronger. And the bibliometric half is not redundant with Kuskova et al. or Wahle et al.: the former analyzed one journal in one field with no outcome variable, the latter aggregated twenty fields without sub-community resolution and never touched AMR; we reuse both methods at a granularity neither attempted, tied to a concrete downstream failure.
Decisive Checks
The proposal stands or falls on three numbers. If $S_i(t)$ does not rise under volume adjustment, the silos are an artifact of publication growth and the framing dies in Experiment 1. If curation lag does not track sub-field distance, the consequence claim dies in Experiment 2 — and we would report that honestly, because it would mean the curation bottleneck (SCIARG's motivation) has some other cause. And if boundary-aware conditioning cannot beat the re-run 77.0% F-score ensemble while narrowing the sub-field spread, the diagnosis may be right but the fix wrong, leaving the bibliometric result as the contribution and the intervention as a documented negative result — still worth having, since Kuskova et al. showed that over-engineered methods reduce reach, and knowing where the cheap fix fails is exactly the kind of constraint the next proposal needs.
References
- Dissemination of Verona Integron-encoded Metallo-β-lactamase among clinical and environmental Enterobacteriaceae isolates in Ontario, Canada
- Data Mining: An AI Perspective
- Visualization as Seen through its Research Paper Keywords
- Towards Automatic Curation of Antibiotic Resistance Genes via Statement Extraction from Scientific Papers: A Benchmark Dataset and Models
- Accelerating Antibiotic Discovery with Large Language Models and Knowledge Graphs
- Citation Amnesia: On The Recency Bias of NLP and Other Academic Fields
- Nanoscience and technology publications and patents: a review of social science studies and search strategies
- Expanding the boundaries of interdisciplinary field: Contribution of Network Science journal to the development of network science
- Mitigating Boundary Ambiguity and Inherent Bias for Text Classification in the Era of Large Language Models
- Optimizing Gene-Based Testing for Antibiotic Resistance Prediction