Turning Citation Networks Inside Out: Studying Science Using Content-Based Knowledge Graphs from LLM-Derived Taxonomies

TL;DR

This study replaces traditional citation-based mappings with a content-driven knowledge graph built from LLM-derived taxonomies, encoding each paper as a triplet of conceptual components $(\mathcal{M}, \mathcal{D}, \mathcal{R})$. Edges capture co-occurrence within six time periods, forming a multi-period tripartite graph analyzed through centrality, co-occurrence, and capacitance metrics to reveal the field’s structural backbone and temporal rearrangements. The findings show a persistent backbone around regression-based measures, with node-level centrality being relatively stable while pairwise and triadic patterns vary, and identify high-capacitance bridges that signal underexplored but potentially important knowledge connections. The approach demonstrates that content-level graphs can uncover meaningful structure and evolution in scientific fields, offering a complementary lens to citation networks with potential broad applicability across domains.

Abstract

Scientific fields are often mapped using citations and metadata, despite knowledge being transmitted primarily through content. We introduce an 'inside-out' approach that reconstructs field structure directly from text by representing each paper as a small set of interpretable knowledge components. Using a large language model to induce domain-specific taxonomies and label papers, each publication is encoded as a triplet of measure, data type, and research-question type. These triplets define a knowledge graph with edges weighted by shared papers. Applied to 617 studies on intergenerational wealth mobility, the graph reveals a stable methodological backbone centered on regression-based mobility measures, alongside substantial temporal variation in component recombination. We further utilize normalized betweenness-to-connectivity ratios to identify components and pairings that act as structural bridges disproportionate to their prevalence. This content-derived, taxonomy-driven mapping complements citation-based approaches by exposing the evolving architecture of methods, data, and questions that define a field.

Figures (11)

• Figure 1: Framework overview.a. From an initial corpus of 16,819 publications, LLM-assisted curation identifies 617 relevant studies on intergenerational mobility. b. Using LLM-driven classification, each paper is assigned to three taxonomies -- measures, data types, and research-question types -- derived from the LLM-generated category sets. c. These taxonomies form the basis of a knowledge graph where edges represent co-occurrence between categories, weighted by the number of shared papers, and the knowledge graph is analyzed with centrality measures.
• Figure 2: Popular nodes, pairs, and triangles in the intergenerational wealth mobility literature across six periods. Each panel summarizes the distribution of components per period. Highlighted markers denote the five most frequent components, presented in the order of frequency. The black dashed line indicates Kendall's $\tau_B$ between two consecutive periods, computed over the common set of components. A 95% bootstrap CI based on 1000 resamples is included to assess robustness given the small sample sizes per period. $\langle \tau \rangle$ denotes the mean observed coefficient across periods, with its 95% bootstrap CI reported alongside. a. Node occurrences (strength/2). Between T3-T5, highlighted nodes dominate, with an increase in $\mathcal{M}_1$ and $\mathcal{R}_4$ at T3. At T6, the three gray nodes between $\mathcal{D}_1$ and $\mathcal{D}_{13}$ correspond to $\mathcal{D}_6, \mathcal{D}_2,$ and $\mathcal{M}_2$. Kendall's $\tau_B$ ranges between $0.64$ and $0.83$, indicating relatively high similarity between consecutive periods, and this similarity stays stable as reflected by the narrow bootstrap CIs (all above 0.5). b. Pairwise co-occurrences. While $\mathcal{M}_1\text{-}\mathcal{R}_2$ and $\mathcal{M}_1\text{-}\mathcal{D}_1$ peak at T4 and decline thereafter, $\mathcal{M}_1\text{-}\mathcal{R}_2$ continues to increase and become dominant at T5. Kendall's $\tau_B$ is lowest between the first two periods ($0.23$) but stabilizes thereafter, reaching a maximum of $0.64$. Rank stability is low in the early comparison, and, the subsequent values remain above zero. c. Three-way co-occurrences.$\mathcal{M}_1\text{-}\mathcal{D}_1\text{-}\mathcal{R}_4$ peaks at T4 before falling below 10 observations by T6, allowing $\mathcal{M}_1\text{-}\mathcal{D}_1\text{-}\mathcal{R}_2$ to emerge as the dominant triangle from T5 onward. The first two periods exhibit minimal rank similarity ($\tau_B=0.04$). Similarity increases over time, albeit insignificantly, reaching a maximum of $\tau_B=0.5$, before slightly declining between the last comparison. Overall, stability remains low, particularly in the first three period-to-period transitions.
• Figure 3: Structurally central nodes and pairs in the intergenerational wealth mobility literature. Each panel shows weighted betweenness centrality across periods. The same color scheme and Kendall's $\tau$ in the black dashed line as in Fig. \ref{['fig:occurrences']} are used. Moreover, the same 95% bootstrap CI and the mean observed coefficient across periods, $\langle \tau \rangle$, are calculated. a. Weighted Node Betweenness.$\mathcal{M}_1$ emerges as the most prominent bridging node from T3 onwards, suggesting that most papers have employed Regression-based Measures ($\mathcal{M}_1$). Although highlighted, $\mathcal{D}_1$ exhibits lower betweenness than $\mathcal{R}_1$ (Measurement and Methodological Advances) at T3, and both $\mathcal{D}_1$ and $\mathcal{D}_{13}$ fall below $\mathcal{R}_3$ (Policy, Institutional, and Geographic Impacts) at T5. This indicates that high-frequency nodes do not necessarily occupy structurally diverse positions. Kendall's $\tau_B$ peaks at the comparison between T2 and T3 (0.75), despite the overlapping bootstrap CIs. All values remain above zero. b. Weighted Edge betweenness. The edge betweenness scores are generally lower than node betweenness, and the highlighted ones are no longer distinct from the gray ones. This suggests that pair-wise connectivity is less diversified, and that frequently occurring pairs do not always correspond to structurally central edges. At T1 and T2, $\mathcal{D}_{13}\text{-}\mathcal{R}_4$ (No dataset-Intergenerational Wealth Mobility and Inheritance) in gray attains the highest betweenness. At T3, $\mathcal{M}_1\text{-}\mathcal{D}_{13}$, records the second highest after $\mathcal{M}_1\text{-}\mathcal{R}_4$; at T5, $\mathcal{M}_1\text{-}\mathcal{R}_3$ follows the highlighted, $\mathcal{M}_1\text{-}\mathcal{R}_4$; and at T6, it follows $\mathcal{M}_1\text{-}\mathcal{R}_2$. $\mathcal{M}_1\text{-}\mathcal{R}_2$ (Regression-based Measures-Empirical Estimates and Determinants) increases steadily and becomes the highest at T5, whereas other periods are dominated by $\mathcal{M}_1\text{-}\mathcal{R}_4$ (Regression-based Measures-Intergenerational Wealth Mobility and Inheritance). The rank similarity between T3 and T4 appears to be highest at the median.
• Figure 4: Median betweenness-degree/count ratios. Each point shows the period-median ratio for a node or pair. With the same color codes as above, the error bars are the range (min-max) of the ratio across six periods. Those that are not highlighted but have their ratio in the 95th percentile are colored in light blue. a. Unweighted Node Capacitance (Eq. \ref{['eq:ratio_degree']}). Given that weights are ignored in calculating degrees, unweighted node betweenness is utilized for the ratio. The highlighted ones, which are the five most occurring nodes in the literature, not only have a high degree but also have a high unweighted node capacitance. b. Weighted Node Capacitance (Eq. \ref{['eq:ratio_strength']}). Given that weights are considered in calculating strengths, weighted node betweenness is used for the ratio. Among the highlighted nodes, $\mathcal{D}_{13}$ is less frequently used in the literature than $\mathcal{D}_1$ but has a higher weighted node capacitance. c. Unweighted Edge Capacitance (Eq. \ref{['eq:ratio_pair']}). Note that some edges only appear once in the network, hence no error bar. We consider the unweighted edge betweenness and edge degree. At the edge level, the five most popularly used elements have low median efficiency, unlike the other two panels. The highest capacitance on median is performed by $\mathcal{D}_{13}\text{-}\mathcal{R}_5$ (No dataset--Demographic Differences in Mobility (Race, Gender, etc.)).
• Figure 5: Representing academic papers at the content-level summarizing all periods. The knowledge graph is constructed by representing each of 617 papers as a triplet -- i.e., each paper consists of a measure, a data type, and a research question type. For example, Bail et al. (2021) paper is represented as $\mathcal{M}_1$-$\mathcal{D}_1$-$\mathcal{R}_4$ and Chetty et al. (2014) paper as $\mathcal{M}_2$-$\mathcal{D}_6$-$\mathcal{R}_2$. Notice that although the used data structure is a multi-period multigraph, this visualization summarizes it across all six periods as a simple network.