Mapping Cultural Connectivity: Network and Embeddings Analysis Applied to English Wikipedia’s Internal Link Structure

PhD Program

Speaker

Paschalis Agapitos

When

2026/10/16
11:00

Place

Salón de Grados - Department of Philosophy (UPV/EHU), Donostia / San Sebastián

Add to calendar

iCal

PhD Thesis defense by Paschalis Agapitos

Supervisor: Gustavo Ariel Schwartz (DIPC Associate Researcher) and Juan Luis Suárez

Computational Humanities

Cultural relationships between entities often leave no direct trace in the historical record, surfacing instead as diffuse patterns of association across large bodies of data. This thesis develops a computational approach, combining network analysis and distributional semantics with humanistic interpretation, to recover such patterns from Wikipedia's link network and examine how they are encoded within large-scale cultural data.

Chapter 1 establishes the conceptual and methodological foundations of the thesis. It opens by framing the core problem: many culturally meaningful relationships are not recorded as direct links between two entities but emerge indirectly, as patterns of association across large bodies of data that traditional close reading was never designed to handle. The chapter then introduces how this thesis addresses this problem by using "focus reading" that combines computational methods, network analysis and distributional semantics, to detect patterns at scale with humanistic frameworks to interpret their significance, situating this approach within digital humanities and cultural data analytics. It explains why the network is treated as a deliberate epistemological choice rather than a neutral mirror of reality, and defines what distinguishes a specifically cultural network: one in which relationships between people are mediated by shared objects, ideas, and institutions. The chapter argues that Wikipedia's network constitutes a suitable, if normatively filtered, instance of such a network, and closes by motivating the study and presenting the three research questions that structure the remaining chapters.

Chapter 2 presents the methodology shared across the three studies and the specific techniques each requires. It explains that all three draw on the same analytical toolkit: network science, which examines the structure of systems of connected entities, and distributional semantics, which represents meaning through context, with each study applying these tools at a different scale. For the seventeenth-century study, the chapter introduces a measure of structural relatedness inspired by the Normalised Google Distance, which infers conceptual closeness between articles from shared linking patterns rather than from direct links alone, and pairs this measure with bootstrapping to test whether observed differences are statistically robust. It then introduces WikiTextGraph, the open-source tool developed for the data mining phase of this thesis, describing how it parses Wikipedia dumps efficiently and reconstructs the internal link network in a way that is reproducible and accessible to non-specialists. Finally, it outlines the procedure for the large-scale biographical study, including the use of Wikidata to reliably detect nearly two million biographies and the combination of network topology with semantic embeddings to recover long-range patterns across twenty-five centuries.

Chapter 3 presents the results, divided into two parts corresponding to the two studies conducted.
Section 3.1 reports the results of the first study, which serves as a proof of concept for the overall approach. It asks whether wikilink-based network analysis can recover historically meaningful patterns within a defined period, testing this on the cultural network connecting Art, Science, and Philosophy in seventeenth-century Europe. Rather than analysing a single triad of figures, as earlier studies had done, it averages across 465 triads sampled from the period, so that the results describe the period itself rather than any one configuration of individuals. The analysis operates at three scales. At the level of the whole network, high modularity shows that the period divides cleanly into three disciplinary clusters. At the cluster level, the measures reveal that Philosophy was the most internally cohesive discipline and that the connection between Science and Philosophy was by far the strongest, while Art occupied a peripheral position, findings that align with what historians have independently established about the period. At the level of individual nodes, the study identifies two distinct statistical patterns of connectivity, a core-periphery structure within disciplines and a heavy-tailed distribution across them and recovers central figures and concepts that were never chosen as starting points. Together, these results confirm that Wikipedia's network encodes recoverable cultural structure and that the method is sensitive to genuine historical change.

Section 3.2 reports the large-scale study, which extends the approach from a single period to roughly 1.9 million biographies spanning twenty-five centuries, combining network analysis with semantic embeddings. The network analysis shows that biographies mostly link to others born within the same or adjacent century, a pattern of temporal homophily so strong that the average gap between linked individuals, about thirty-four years, closely matches the span of a human generation. The sparse long-range links that remain are not random: connections from later scientists back to Classical figures trace the documented transmission of ancient knowledge, while the choice of how links are counted determines whether the network tells a story of historical continuity or of rupture, a reminder that such methodological choices are interpretive rather than neutral. The semantic analysis then finds that centuries close in time occupy nearby regions of the embedding space, conventional historical periods re-emerge from the data without being imposed, and semantic similarity declines gradually with temporal distance, a regularity termed here the temporal Tobler pattern. The outliers to that trend surface periods that are distant in time yet conceptually close, such as the link between thirteenth-century science and Classical antiquity through the Arabic-Latin translation movement. Taken together, the section shows that both the network and the semantic structure of Wikipedia's biographical network are organised by time, and that the patterns recoverable from that network reflect not only historical processes but the representational choices (methodological and editorial) through which those processes have been made legible at scale.

This thesis was carried out by the author at the Centro de Física de Materiales (CFM) under the supervision of Dr. Gustavo Ariel Schwartz (CFM) and Dr. Juan Luis Suárez (CulturePlex Lab, London, Ontario, Canada), with funding provided by the Donostia International Physics Centre (DIPC) through the Mestizajes project.