Appendix A1. Operationalisation of Indicators and Statistics

Panel A. Knowledge Production Mode (5 indicators, 23 statistics)

Indicator	Statistic	Formula	Symbol explanation	Source
Citation behaviour	citation density	$D_{c} = \frac{N_{cite}}{N_{word}} \times 1000$	D_c = citation density; N_cite = number of in-text citation markers; N_word = number of non-punctuation words	Fowler & Aksnes (2007)
Citation behaviour	citation age	$A = y_{paper} - \frac{1}{N_{ref}} \sum_{i = 1}^{N_{ref}} y_{i}$	A = citation age; y_paper = publication year of the citing paper; y_i = publication year of reference i; N_ref = number of references	Fowler & Aksnes (2007)
Citation behaviour	self-citation rate	$R_{s} = \frac{N_{self}}{N_{ref}}$	R_s = self-citation rate; N_self = references sharing at least one author surname with the citing paper; N_ref = total number of references	Fowler & Aksnes (2007)
Lexical diversity	MTLD	$MTLD = \frac{1}{2} ({MTLD}_{fwd} + {MTLD}_{bwd})$	MTLD_fwd = forward MTLD; MTLD_bwd = backward MTLD; N = total number of tokens; F = number of factors until TTR ≤ 0.72	McCarthy & Jarvis (2010)
Lexical diversity	Simpson’s D	$D = \sum_{i = 1}^{V} {(\frac{n_{i}}{N})}^{2}$	D = Simpson’s D; V = number of distinct word types; n_i = frequency of word type i; N = total tokens	Simpson (1949)
Lexical diversity	Simpson’s Diversity	$1 - D$	Simpson’s Diversity = 1 − D; D = Simpson’s D as defined above	Simpson (1949)
Lexical diversity	MATTR	$MATTR = \frac{1}{N - W + 1} \sum_{i = 1}^{N - W + 1} \frac{\| V_{i} \|}{W}$	MATTR = moving-average type-token ratio; W = 50 (window length); N = total tokens; \|V_i\| = number of distinct types in window i	Covington & McFall (2010)
Syntactic complexity	mean dependency distance (MDD)	$MDD = \frac{1}{N_{dep}} \sum_{i = 1}^{N_{dep}} \| i - head (i) \|$	MDD = mean dependency distance; N_dep = number of non-root tokens; i = linear position of a token; head(i) = linear position of its syntactic head	Liu (2008)
Syntactic complexity	mean hierarchical distance (MHD)	$MHD = \frac{1}{N_{dep}} \sum_{i = 1}^{N_{dep}} d (i)$	MHD = mean hierarchical distance; d(i) = number of edges from token i to the root; N_dep = number of non-root tokens	Jing & Liu (2015)
Syntactic complexity	maximum dependency distance	${max}_{i} \| i - head (i) \|$	Maximum over all non-root tokens i of \|i − head(i)\|; head(i) = linear position of its syntactic head	Jiang & Jiang (2020)
Syntactic complexity	maximum tree depth	${max}_{i} d (i)$	Maximum over all non-root tokens i of d(i); d(i) = number of edges from token i to the root	Jiang & Liu (2015)
Syntactic complexity	average sentence length	$\frac{N_{word}}{N_{sent}}$	N_word = number of non-punctuation tokens; N_sent = number of sentences	—
Syntactic complexity	number of sentences	$N_{sent}$	N_sent = number of sentences in the paper	—
Punctuation density	comma per 1,000 words	$\frac{N_{comma}}{N_{word}} \times 1000$	N_comma = number of commas; N_word = number of non-punctuation tokens	—
Punctuation density	semicolon per 1,000 words	$\frac{N_{semi}}{N_{word}} \times 1000$	N_semi = number of semicolons; N_word = number of non-punctuation tokens	—
Punctuation density	parenthesis per 1,000 words	$\frac{N_{paren}}{N_{word}} \times 1000$	N_paren = number of parentheses; N_word = number of non-punctuation tokens	—
Punctuation density	dash per 1,000 words	$\frac{N_{dash}}{N_{word}} \times 1000$	N_dash = number of dashes; N_word = number of non-punctuation tokens	—
Section structure	number of sections	$\| {s : s is a non-empty section label} \|$	Number of distinct non-empty section labels s in the main body text	—
Section structure	Introduction ratio	$\frac{W_{intro}}{W_{total}}$	W_intro = number of words in sections labeled “Introduction”; W_total = total words across all sections	—
Section structure	Related Work ratio	$\frac{W_{related}}{W_{total}}$	W_related = number of words in sections labeled “Related Work” or “Background”; W_total = total words across all sections	—
Section structure	Method ratio	$\frac{W_{method}}{W_{total}}$	W_method = number of words in sections labeled “Method”, “Approach”, or “Model”; W_total = total words across all sections	—
Section structure	Experiments ratio	$\frac{W_{exp}}{W_{total}}$	W_exp = number of words in sections labeled “Experiments”, “Evaluation”, or “Results”; W_total = total words across all sections	—
Section structure	Conclusion ratio	$\frac{W_{concl}}{W_{total}}$	W_concl = number of words in sections labeled “Conclusion” or “Discussion”; W_total = total words across all sections	—

Panel B. Knowledge Producers (3 indicators, 7 statistics)

Indicator	Statistic	Formula	Symbol explanation	Source
Author scale and collaboration structure	average number of authors per paper	$\bar{a} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} a_{j}$	a_j = number of authors of paper j; N_p = number of papers in the year or stage	Wuchty et al. (2007)
Author scale and collaboration structure	proportion of single-authored papers	$p_{single} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I (a_{j} = 1)$	p_single = proportion of single-authored papers; a_j = number of authors of paper j; I(.) = indicator function (1 if true, 0 otherwise); N_p = number of papers	—
Author scale and collaboration structure	Gini coefficient of author team size	$G = \frac{n + 1 - 2 \sum_{i = 1}^{n} c_{i} / c_{n}}{n}$	G = Gini coefficient; c_i = cumulative sum of sorted author counts; c_n = total sum of author counts; n = number of papers	Gini (1912)
Venue ecology	venue diversity entropy	$H_{v} = - \sum_{v = 1}^{V} p_{v} \log_{2} p_{v}$	H_v = venue diversity entropy; V = number of distinct venues; p_v = proportion of papers in venue v	Shannon (1948)
Venue ecology	new venue emergence rate	$r_{new} = \frac{N_{new}}{N_{venue}}$	r_new = new venue emergence rate; N_new = number of venues appearing for the first time in the given year; N_venue = total active venues in that year	—
Venue ecology	number of venues	$N_{venue} = \| {v : v is a non-empty venue} \|$	N_venue = number of distinct non-empty venue labels v in the year or stage	—
SIG coverage	SIG coverage rate	$p_{SIG} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I ({SIGs}_{j} \neq \emptyset)$	p_SIG = SIG coverage rate; SIGs_j = SIG field of paper j; I(.) = indicator function; N_p = number of papers	—

Note. Venue survival duration, defined as the number of years between a venue’s first and last appearance, is computed in the analysis pipeline as a supplementary descriptive statistic. It is not included among the 76 statistics or in the stage-wise significance tests. Inter-institutional collaboration rate could not be computed because the metadata contain no author affiliation information.

Panel C. Knowledge Evaluation (3 indicators, 6 statistics)

Indicator	Statistic	Formula	Symbol explanation	Source
Openness and reproducibility	attachment supply rate	$p_{art} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I ({Att}_{j} \neq \emptyset \lor {Data}_{j} \neq \emptyset \lor {Soft}_{j} \neq \emptyset)$	p_art = attachment supply rate; Att_j, Data_j, Soft_j = Attachment, Dataset, Software fields of paper j; I(.) = indicator function; N_p = number of papers	—
Openness and reproducibility	reproducibility premium	$d = \frac{{\bar{x}}_{art} - {\bar{x}}_{no-art}}{s_{pooled}}$	d = Cohen’s d; x̄_art = mean in-degree of papers with artifacts; x̄_no-art = mean in-degree of papers without artifacts; s_pooled = pooled standard deviation	Cohen (1988)
Self-correction and iteration	version iteration rate	$p_{ver} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I ({Ver2}_{j} \neq \emptyset \lor {Ver3}_{j} \neq \emptyset \lor {Ver4}_{j} \neq \emptyset)$	p_ver = version iteration rate; Ver2_j, Ver3_j, Ver4_j = Version 2, 3, 4 fields; I(.) = indicator function; N_p = number of papers	—
Self-correction and iteration	erratum rate	$p_{err} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I ({Erratum}_{j} \neq \emptyset)$	p_err = erratum rate; Erratum_j = Erratum field of paper j; I(.) = indicator function; N_p = number of papers	—
Recognition and reward	award distribution	$p_{award} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I ({Award}_{j} \neq \emptyset)$	p_award = award distribution; Award_j = Award field of paper j; I(.) = indicator function; N_p = number of papers	—
Recognition and reward	award type entropy	$H_{award} = - \sum_{a = 1}^{A} p_{a} \log_{2} p_{a}$	H_award = award type entropy; A = number of distinct award types; p_a = proportion of award type a among awarded papers	Shannon (1948)

Panel D. Knowledge Content (11 indicators, 29 statistics)

Indicator	Statistic	Formula	Symbol explanation	Source
SND — Target-Neighbour similarity (TN)	TN	$TN (w) = \frac{1}{k} \sum_{i = 1}^{k} \cos (𝐱_{w}, 𝐱_{n_{i} (w)})$	TN(w) = Target-Neighbour similarity; w = target word; x_w = type-level mean embedding; n_i(w) = i-th nearest neighbour; k = 10	Schulte im Walde & Frassinelli (2022)
SND — Target-Neighbour similarity (TN)	ΔND(TN) 1→2	$Δ {ND}_{1 \to 2} (w) = {TN}_{2} (w) - {TN}_{1} (w)$	ΔND(TN) = difference in TN between Stage 2 and Stage 1	Miletic & Falk (2026)
SND — Target-Neighbour similarity (TN)	ΔND(TN) 2→3	$Δ {ND}_{2 \to 3} (w) = {TN}_{3} (w) - {TN}_{2} (w)$	ΔND(TN) = difference in TN between Stage 3 and Stage 2	Miletic & Falk (2026)
SND — Target-Neighbour similarity (TN)	ΔND(TN) 3→4	$Δ {ND}_{3 \to 4} (w) = {TN}_{4} (w) - {TN}_{3} (w)$	ΔND(TN) = difference in TN between Stage 4 and Stage 3	Miletic & Falk (2026)
SND — Target-Neighbour similarity (TN)	ΔND(TN) 1→4	$Δ {ND}_{1 \to 4} (w) = {TN}_{4} (w) - {TN}_{1} (w)$	ΔND(TN) = difference in TN between Stage 4 and Stage 1	Miletic & Falk (2026)
SND — Neighbour-Neighbour similarity (NN)	NN	$NN (w) = \frac{1}{(\binom{k}{2})} \sum_{i = 1}^{k} \sum_{j = i + 1}^{k} \cos (𝐱_{n_{i} (w)}, 𝐱_{n_{j} (w)})$	NN(w) = Neighbour-Neighbour similarity; n_i(w) and n_j(w) = k nearest neighbours (k = 10); binom(k,2) = number of neighbour pairs	Schulte im Walde & Frassinelli (2022)
SND — Neighbour-Neighbour similarity (NN)	ΔND(NN) 1→2	$Δ {ND}_{1 \to 2} (w) = {NN}_{2} (w) - {NN}_{1} (w)$	ΔND(NN) = difference in NN between Stage 2 and Stage 1	Miletic & Falk (2026)
SND — Neighbour-Neighbour similarity (NN)	ΔND(NN) 2→3	$Δ {ND}_{2 \to 3} (w) = {NN}_{3} (w) - {NN}_{2} (w)$	ΔND(NN) = difference in NN between Stage 3 and Stage 2	Miletic & Falk (2026)
SND — Neighbour-Neighbour similarity (NN)	ΔND(NN) 3→4	$Δ {ND}_{3 \to 4} (w) = {NN}_{4} (w) - {NN}_{3} (w)$	ΔND(NN) = difference in NN between Stage 4 and Stage 3	Miletic & Falk (2026)
SND — Neighbour-Neighbour similarity (NN)	ΔND(NN) 1→4	$Δ {ND}_{1 \to 4} (w) = {NN}_{4} (w) - {NN}_{1} (w)$	ΔND(NN) = difference in NN between Stage 4 and Stage 1	Miletic & Falk (2026)
SND — Average Radius of Co-occurrence (ARC)	ARC	$ARC (w) = \frac{1}{\| N_{s} (w) \|} \sum_{y \in N_{s} (w)} \cos (𝐱_{w}, 𝐱_{y})$	ARC(w) = Average Radius of Co-occurrence; N_s(w) = {y : d(x_w, x_y) ≤ τ}; τ = μ_d − 1.5σ_d	Shaoul & Westbury (2010)
SND — Average Radius of Co-occurrence (ARC)	ΔND(ARC) 1→2	$Δ {ND}_{1 \to 2} (w) = {ARC}_{2} (w) - {ARC}_{1} (w)$	ΔND(ARC) = difference in ARC between Stage 2 and Stage 1	Miletic & Falk (2026)
SND — Average Radius of Co-occurrence (ARC)	ΔND(ARC) 2→3	$Δ {ND}_{2 \to 3} (w) = {ARC}_{3} (w) - {ARC}_{2} (w)$	ΔND(ARC) = difference in ARC between Stage 3 and Stage 2	Miletic & Falk (2026)
SND — Average Radius of Co-occurrence (ARC)	ΔND(ARC) 3→4	$Δ {ND}_{3 \to 4} (w) = {ARC}_{4} (w) - {ARC}_{3} (w)$	ΔND(ARC) = difference in ARC between Stage 4 and Stage 3	Miletic & Falk (2026)
SND — Average Radius of Co-occurrence (ARC)	ΔND(ARC) 1→4	$Δ {ND}_{1 \to 4} (w) = {ARC}_{4} (w) - {ARC}_{1} (w)$	ΔND(ARC) = difference in ARC between Stage 4 and Stage 1	Miletic & Falk (2026)
SND — Token-level SND	token-level SND	${SND}_{tok} (w) = \frac{1}{(\binom{n}{2})} \sum_{i = 1}^{n} \sum_{j = i + 1}^{n} \cos (𝐞_{i}, 𝐞_{j})$	SND_tok(w) = token-level SND; e_i = contextualised embedding of occurrence i; n = total occurrences of w	Sagi et al. (2009)
SND — Token-level SND	ΔND(token-SND) 1→2	$Δ {ND}_{1 \to 2} (w) = {SND}_{tok, 2} (w) - {SND}_{tok, 1} (w)$	ΔND(token-SND) = difference in token-level SND between Stage 2 and Stage 1	Miletic & Falk (2026)
SND — Token-level SND	ΔND(token-SND) 2→3	$Δ {ND}_{2 \to 3} (w) = {SND}_{tok, 3} (w) - {SND}_{tok, 2} (w)$	ΔND(token-SND) = difference in token-level SND between Stage 3 and Stage 2	Miletic & Falk (2026)
SND — Token-level SND	ΔND(token-SND) 3→4	$Δ {ND}_{3 \to 4} (w) = {SND}_{tok, 4} (w) - {SND}_{tok, 3} (w)$	ΔND(token-SND) = difference in token-level SND between Stage 4 and Stage 3	Miletic & Falk (2026)
SND — Token-level SND	ΔND(token-SND) 1→4	$Δ {ND}_{1 \to 4} (w) = {SND}_{tok, 4} (w) - {SND}_{tok, 1} (w)$	ΔND(token-SND) = difference in token-level SND between Stage 4 and Stage 1	Miletic & Falk (2026)
Semantic drift — PRT	prototypical embedding distance (PRT)	$PRT = 1 - \cos (\frac{1}{N_{t 1}} \sum_{i \in t 1} 𝐞_{i}, \frac{1}{N_{t 2}} \sum_{j \in t 2} 𝐞_{j})$	PRT = prototypical embedding distance; t1, t2 = two time periods; N_t1, N_t2 = occurrence counts in each period	Kutuzov & Giulianelli (2020)
Semantic drift — APD	average pairwise distance (APD)	$APD = 1 - \frac{1}{N_{t 1} N_{t 2}} \sum_{i \in t 1} \sum_{j \in t 2} \cos (𝐞_{i}, 𝐞_{j})$	APD = average pairwise distance; t1, t2 = two time periods; N_t1, N_t2 = occurrence counts in each period	Giulianelli et al. (2020)
Semantic drift — composite	composite Drift	$Drift = \frac{PRT + APD}{2}$	Composite Drift = arithmetic mean of PRT and APD	Kutuzov & Giulianelli (2020)
Term lifecycle	term half-life	$T_{half} = y^{*} - y_{peak}$	T_half = term half-life; y_peak = year of peak frequency; y^* = first subsequent year in which frequency drops to half	Burton & Kebler (1960)
Term lifecycle	new terms per year	$N_{new} (y) = \| {t : first year of t = y} \|$	N_new(y) = number of terms t appearing for the first time in year y	—
Term lifecycle	unique terms per year	$N_{unique} (y) = \| {t : f_{t} (y) > 0} \|$	N_unique(y) = number of distinct terms t in year y; f_t(y) = frequency of term t in year y	—
Language coverage entropy	language coverage entropy	$H_{lang} (y) = - \sum_{l = 1}^{L_{y}} p_{l} \log_{2} p_{l}$	H_lang(y) = language coverage entropy; L_y = number of languages in year y; p_l = proportion of language l mentions in year y	Shannon (1948)
Low-resource language ratio	low-resource language ratio	$p_{low} (y) = \frac{1}{N_{y}^{valid}} \sum_{j = 1}^{N_{y}^{valid}} I ({lang}_{j} \cap ℒ_{low} \neq \emptyset)$	p_low(y) = low-resource language ratio; N_y^valid = papers with language info in year y; lang_j = languages of paper j; L_low = bottom 50% languages; I(.) = indicator	Joshi et al. (2020)
Multilingual paper ratio	multilingual paper ratio	$p_{multi} (y) = \frac{1}{N_{y}} \sum_{j = 1}^{N_{y}} I ({title}_{j} or {abstract}_{j} matches 𝒦)$	p_multi(y) = multilingual paper ratio; N_y = papers in year y; K = keyword set; I(.) = indicator	—

Note. TN, NN, ARC, and token-level SND are computed for each target word and each stage; ΔND is computed for all adjacent-stage pairs and for the first–last pair. PRT, APD, and composite Drift are computed at the annual, stage, and stage-transition levels; the annual series are used in RQ2 change-point detection, and the stage-level means enter RQ1 significance testing. Embeddings are extracted with ModernBERT-base, and the last four hidden layers are averaged dimension-wise. PCA retains the top 10% of principal components per target word before drift computation.

Panel E. Knowledge Packaging (6 indicators, 11 statistics)

Indicator	Statistic	Formula	Symbol explanation	Source
Title length	title length in words	$L_{words} = \| title.split() \|$	L_words = number of whitespace-delimited tokens in the title	—
Title length	title length in characters	$L_{chars} = \| title \|$	L_chars = number of characters in the title	—
Title colon rate	title colon rate	$p_{colon} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I (':' \in {title}_{j})$	p_colon = title colon rate; I(.) = indicator function (1 if true, 0 otherwise)	—
Title question rate	title question rate	$p_{question} = \frac{1}{N_{p}} \sum_{j = 1}^{N_{p}} I ('?' \in {title}_{j})$	p_question = title question rate; I(.) = indicator function	—
Title NP complexity	NP mean length	${\bar{L}}_{NP} = \frac{1}{\| 𝒩 \|} \sum_{n \in 𝒩} \| n \|$	L̄_NP = NP mean length; N = set of noun chunks; \|n\| = tokens in chunk n	Berlage (2014); Diao (2021)
Title NP complexity	NP mean dependency count	${\bar{D}}_{NP} = \frac{1}{\| 𝒩 \|} \sum_{n \in 𝒩} \| {t \in n : head (t) = root (n), t \neq root (n)} \|$	D̄_NP = NP mean dependency count; head(t) = head of token t; root(n) = chunk root	Kyle (2016); Biber et al. (2011)
Title NP complexity	NP postmodifier level	${\bar{M}}_{NP} = \frac{1}{\| 𝒩 \|} \sum_{n \in 𝒩} {max}_{t \in n, t . i > root (n) . i} ℓ (dep (t))$	M̄_NP = NP postmodifier level; ℓ(.) = ordinal level (0-3); dep(t) = dependency label	Berlage (2014); Biber et al. (2011)
Title NP complexity	NP sententiality ratio	$p_{sent} = \frac{1}{\| 𝒩 \|} \sum_{n \in 𝒩} I (\exists t \in n : pos (t) \in {VERB, AUX} \lor dep (t) \in 𝒟_{sent})$	p_sent = NP sententiality ratio; pos(.) = part-of-speech; D_sent = clausal deps; I(.) = indicator	Berlage (2014)
Title NP complexity	NP coverage ratio	$C_{NP} = \frac{\sum_{n \in 𝒩} \| n \|}{\| {t \in title : \neg is\_punct (t)} \|}$	C_NP = NP coverage ratio; numerator = total NP tokens; denominator = non-punct title tokens	Diao (2021)
Figure/table type ratio	figure/table type ratio	$r_{fig} = \frac{N_{fig}}{N_{fig} + N_{tab}}$	r_fig = figure/table type ratio; N_fig = figure entries; N_tab = table entries	—
Formula density	formula density	$D_{eq} = \frac{N_{eq}}{N_{word}} \times 1000$	D_eq = formula density; N_eq = equation spans; N_word = total words	—

Note. A composite score of Title noun phrase complexity, defined as the equally weighted mean of the five min–max normalised NP sub-indicators, is computed by the pipeline as a supplementary diagnostic. It is not included among the 11 statistics or in the RQ1 main test summary; it is retained only for internal cross-checking.