Bipartite hop — Linnainmaa 1976's citation histogram vs Nguyen & Eger 2024's AI-bibliography-age-halving: a genuine shared mechanism (exponential field growth + recency skew), tested quantitatively and found to run in the SAME direction, not opposite ones

bipartitebridge-investigationcitation-analysisbibliometricslinnainmaacitation-amnesiafield-growthembedding-bridgearxiv-growth

Bridge-seed check (bridge-seed discipline): searched 10-inbox/raw/ for any capture naming both claim-linnainmaa-1976-citation-histogram and claim-nguyen-eger-2024-ai-bibliography-age-halved-reframes-amnesia-as-dynamism. One file matches "nguyen-eger" (2026-09-25-hop-ml-nlp-measured-insularity-and-garfields-1980-citation-amnesia.md, the capture that originally promoted the Nguyen & Eger claim-note), but it never mentions Linnainmaa — its hop chain runs seed→SciTraj→Wahle→Nguyen & Eger→Garfield, a citation-insularity thread with no contact with the backprop-priority cluster. Three other recovered captures pair the Linnainmaa histogram note against a different bipartite partner (OSTI's SAND-91-7095 metadata record — see observation-osti-semantic-scholar-cosine-pairing-is-verification-genre-false-friend), not against Nguyen & Eger. No sibling capture investigates this specific pair. This is a fresh investigation, not a duplicate.

Claim: Nguyen & Eger's own explanation for the bibliography-age halving is exponential growth in AI subfields' publication volume, not forgetting

claim_type: technical-mechanism (Tier 1 required, cleared — the paper's own text)

Nguyen & Eger do not treat "citation amnesia" as a bias to be corrected; they offer a specific causal mechanism and quantify it. Arguing that NLP is "simply another field like other AI fields which are all characterized by high dynamicity, i.e., many newly incoming researchers (and submissions) and quickly changing state-of-the-art solutions," they measure the growth directly: "cs.CL (Computer Linguistics) has the highest growth rate of almost ≈ 32-times (219 submissions in 2013 to above 7k submissions in 2023), followed by cs.CV (Computer Vision) at ≈ 22-times and cs.LG (Machine Learning) at ≈ 20-times." Figure 7 of the paper plots, per field, "the yearly average citation increases (negative numbers denote decreases) vs. the median yearly submission increases of each field; the latter is an indicator of the dynamicity of the field" — making the submission-growth rate the paper's own proposed explanatory variable for the citation-age trend, not a side observation. Their conclusion states the mechanism plainly: "In such an environment, the observed changes in the age of the bibliography may simply be a 'natural' response."

• source_url: "https://arxiv.org/pdf/2401.03545"
• source_sha: "d3a795de3adcebc114a4a7f519a5698c860c3326ad73a9d68d6b15211736afc9"
• source_title: "Is there really a Citation Age Bias in NLP?"
• source_author: "Hoa Nguyen, Steffen Eger"
• source_date: "2024 (arXiv:2401.03545v1)"
• source_venue: "arXiv cs.DL (author preprint)"
• source_quote: "cs.CL (Computer Linguistics) has the highest growth rate of almost ≈ 32-times (219 submissions in 2013 to above 7k submissions in 2023), followed by cs.CV (Computer Vision) at ≈ 22-times and cs.LG (Machine Learning) at ≈ 20-times"
• source_quote_2: "In such an environment, the observed changes in the age of the bibliography may simply be a 'natural' response."
• source_tier: 1
• Both quotes passed quote_check against the extracted PDF text at capture time.

Claim: Applying the same growth-rate lens to Linnainmaa's own citation histogram shows his paper's citation-rate growth badly underperforming the field's publication-volume growth over a comparable window — the two notes are not independent, and they are not in tension; they are two readings of one process

claim_type: synthesis / arithmetic inference — not a claim made directly by either primary source. The two input figures are each independently Tier 1 (Nguyen & Eger's own submission counts; the vault's Semantic Scholar-sourced Linnainmaa histogram), but their combination here is Seek's calculation, done to actually test the resemblance rather than assert it.

claim-linnainmaa-1976-citation-histogram records 67 citations to Linnainmaa's 1976 paper across the 2010s (≈6.7/year) and 178 across the 2020s through mid-2026, roughly 6.5 elapsed years (≈27.4/year) — a citation-rate increase of ≈4.1× between the two decades. Over a roughly comparable ~10-year window (2013→2022/23), Nguyen & Eger measure AI subfields' arXiv submission volume growing 20–32× (cs.LG ≈20×, cs.CV ≈22×, cs.CL ≈32×).

If the field's exploding paper count were the entire story behind Linnainmaa's absolute citation surge — more papers simply mean more citing events for any still-relevant old work, with no change in per-paper citing behavior — Linnainmaa's citation-rate growth should track the field's submission growth, i.e. land somewhere near 20–32×. It instead lands at roughly an eighth to a quarter of that. In other words: Linnainmaa's paper is being cited more often in absolute terms every decade, but its share of the field's exploding citation volume is shrinking. That shrinking share is exactly the signature Nguyen & Eger's halving-bibliography-age finding predicts for an old paper in a field where citing behavior skews toward recent work even as the field balloons: old, no-longer-the-newest work gets a smaller slice of a much bigger pie. The apparent resemblance between the two seed notes (cosine 0.87, no shared vocabulary) is therefore not a coincidence and not a contradiction — it is the same exponential-growth-plus-recency-skew mechanism observed from two different vantage points: Nguyen & Eger measure it as outgoing-bibliography age across ~300k papers' references; the Linnainmaa histogram is one old paper's incoming-citation trajectory, and the two numbers are consistent with sitting on the same underlying curve.

• source_url: "vault:30-notes/claim-linnainmaa-1976-citation-histogram.md" (per-decade table: 2010s 67, 2020s 178 through mid-2026)
• source_title: "Linnainmaa 1976's citation record, finally pulled — 51 dated citations across 1978–2009, then 245 in the sixteen years since 2010"
• source_tier: 1 (vault-internal primary; underlying data is the Semantic Scholar Graph API, queried 2026-07-07)
• source_url_2: "https://arxiv.org/pdf/2401.03545"
• source_sha_2: "d3a795de3adcebc114a4a7f519a5698c860c3326ad73a9d68d6b15211736afc9"
• source_quote_2: "cs.CL (Computer Linguistics) has the highest growth rate of almost ≈ 32-times (219 submissions in 2013 to above 7k submissions in 2023), followed by cs.CV (Computer Vision) at ≈ 22-times and cs.LG (Machine Learning) at ≈ 20-times"
• source_tier_2: 1
• Arithmetic (4.1× = (178/6.5) / (67/10)) performed by Seek this session; not independently cross-checked against a scripted recount of the Linnainmaa decade counts, which the histogram note itself already flags as a lead.

Claim: The comparison is a cross-dataset extrapolation, not a joint test on matched data — the mechanism is genuine but its exact fit to Linnainmaa is not yet directly verified

claim_type: honest-limits / scope statement

Nguyen & Eger's dataset is ~300k arXiv papers' outgoing references, 2013–2022, measured via Semantic Scholar's reference-extraction; the Linnainmaa histogram is the incoming citation list for one 1976 paper, 1978–2026, measured via Semantic Scholar's citation-graph endpoint. Neither paper is in the other's dataset, the time windows only partially overlap, and nobody has run Linnainmaa through Nguyen & Eger's own metric (mean/median Age-of-Citation, restricted to papers that cite him) or checked whether the field classification he'd fall under (numerical analysis / cs.NA, not cs.AI/cs.CL/cs.CV/cs.LG) is even one Nguyen & Eger measured growth for — it is not in their eight subcategories. The claim above is therefore [inference — could not be fully confirmed with a single matched dataset after this session's search]: the mechanism (field growth plus recency-skewed citing) is real and Tier-1-sourced on the Nguyen & Eger side, and the arithmetic direction (Linnainmaa underperforming field growth) is internally consistent, but a scripted join of the two underlying datasets — which this session did not attempt — would be the actual test.

Safety flags

None fired. The fetched source (arxiv.org/pdf/2401.03545, tls: verified) is a standard academic preprint: third-person scientific prose throughout, no addressed-to-AI language, no override or authority claims, no tier self-assignment, no file-system or credential requests, no urgency framing.

Further leads

• Re-run the Linnainmaa /citations query with fields=title,authors,year, classify the 51 pre-2010 and 245 post-2010 citing papers by field (cs.AI/cs.CL/cs.CV/cs.LG vs numerical-analysis/AD venues), and compute his own Age-of-Citation-style trajectory directly against Nguyen & Eger's metric — the scripted join this capture's third claim says is still missing.
• Singh et al. 2023 ("Forgotten knowledge: Examining the citational amnesia in NLP," ACL 2023) is the paper Nguyen & Eger are explicitly rebutting; not independently fetched this session — the rebuttal was read only through Nguyen & Eger's framing of it.
• Bollmann & Elliott 2020 ("On forgetting to cite older papers: An analysis of the ACL Anthology," ACL 2020) is cited by Nguyen & Eger as the earlier, narrower-window (2010–2019) version of the same finding Singh et al. later confirm and Nguyen & Eger contest — not independently fetched.
• Mukherjee et al. 2017 ("The nearly universal link between the age of past knowledge and tomorrow's breakthroughs," Science Advances) and Parolo et al. 2015 ("Attention decay in science," J. Informetrics), both cited in Nguyen & Eger's related work, bear directly on whether an old paper's citation trajectory is expected to decay or spike — relevant background for testing the Linnainmaa curve specifically, not read this session.
• Verstak et al. 2014 ("On the shoulders of giants: The growing impact of older articles") is cited by Nguyen & Eger for the opposite-direction effect (digital search makes old papers easier to find and cite) — a candidate second mechanism for Linnainmaa's post-2010 surge, not evaluated this session.

Entity candidates

• Janvijay Singh, Mukund Rungta, Saif M. Mohammad, Diyi Yang — persons — authors of Singh et al. 2023, the paper that made the original "citational amnesia in NLP" claim Nguyen & Eger's whole paper is built to contest. Flagged first per the standing blind-spot warning: this is the earlier, foundational work the seed claim's subject (Nguyen & Eger) measures itself against, not a co-author of the rebuttal itself.
• Marcel Bollmann, Desmond Elliott — persons — authors of the 2020 ACL paper that first measured the narrower (2010–2019) version of the same trend; a second foundational-comparison point, same blind spot.
• Seppo Linnainmaa — person — already has an entity page (entity-seppo-linnainmaa); the foundational figure whose single paper's citation trajectory this capture's central arithmetic tests against the field-growth mechanism.
• Hoa Nguyen, Steffen Eger — persons — authors of the seed claim's source paper.
• Exponential field growth (arXiv submission-count multipliers) — concept — the measured variable Nguyen & Eger use to explain bibliography-age dynamics; a strong candidate for its own concept page if a future capture runs the Linnainmaa join and confirms or kills the mechanism directly.