The project started by developing a method to analyze a text corpora using a pretrained language model to extract keyphrases and building keyphrase co-occurrence networks (Jacobetty & Kolleck, 2015). This exploration motivated our decision to explore using text network analysis techniques for analyzing text corpora. We also explored the roles and structures of transnational education networks and the promises/limits of social network analysis for education research (Kolleck, 2015;Schuster, Hansen & Kolleck, 2015). Finally, we analysed discourse-network maps to reveal framing and power asymmetries (Goritz & Kolleck, 2015).
The project progressed through five interlocking work streams that combined computational methods with mixed methods fieldwork. First, a large text corpus (5,425 documents) from ten global education networks was processed with Named Entity Recognition to extract country co-mentions and build weighted country-country discourse networks. Linking these networks to a development metric (Augmented Human Development Index) revealed a pattern in which donor states act as structural hubs while recipient states are frequent objects of mention. This poses a challenge to the “partnership” rhetoric that the network form typically espouses.
Second, the team created a computational framework to map the international education governance literature (4,420 papers), combining dynamic topic modeling with semantic change-point detection. This revealed three thematic areas in the literature and a notable temporal discontinuity around 2009, indicating a post-2009 shift toward more instrumental, audit-oriented research agendas. The method is reusable and intended for broader application across corpora.
Third, a wide organizational survey (≈2,000 organizations contacted; >400 responses; 140 with full network data) revealed new evidence on network structure and outcomes. Key technical findings include: (a) networks seem to differ in whether informational or resource flows dominate; (b) reciprocal information exchange (bidirectional ties) correlates strongly with successful direct service delivery, while resource ties are more associated with legal/institutional change; and (c) many organisations feel structurally dependent despite valuing connections.
Fourth, qualitative work (semi-structured interviews; regional case study of civic education in Europe with 583 organisations) supported triangulation: interviews illuminated motivations for networking (information access, visibility, credibility, protection), while the European mapping exposed fragmentation, short-term funding pressures, and the link between funding diversity and innovation.
Finally, the team developed two open-source data-collection frameworks (TikTok Research API and Wikidata API) and an text analysis pipeline (topic model + change-point detection). These tools further enable the project’s large-scale textual and network extraction, and will be publicly shared with documentation. Together, the methods, datasets and code form our computational education research’s methodological infrastructure.
