This study, which forms an integral part of a broader investigation into the interconnectedness of educational discourses, focuses on perspectives of knowledge. These perspectives encompass the epistemological frameworks of knowledge, including its nature, acquisition, validation, and utilisation, and are influenced by various schools of thought, such as empiricism, pragmatism, and constructivism. These perspectives are pivotal not only to our perception of knowledge but also have a profound impact on teaching methods (Greene & Yu, 2016) and learner outcomes (Mason et al., 2013; Muis & Foy, 2010).

The project examines the last 60 years of educational discourse in Sweden (1962–2023). By analysing extensive textual data from parliamentary records, media, and educational research using innovative digital methods, we aim to illustrate how perspectives on knowledge have evolved within and across different discourse domains.

We chose 1962 as the starting point for our study because it marks the introduction of Sweden's comprehensive educational system. Since then, Sweden has experienced significant transformations in its educational system, including a shift from state to municipal governance of schools, the introduction of privately run schools, the establishment of a new teaching college, and the implementation of a new grading system (Lindensjö & Lundgren, 2000). These reforms align with major developments in Sweden's economic history, such as comprehensive welfare reforms and sustained economic growth (Schön, 2012).

Presently, the Swedish school system is grappling with a multitude of challenges, including declining academic achievement, increased inequality, grade inflation, inadequate competence supply, classroom disorder, and mental health issues. Recent research suggests that a transformed perception of knowledge underlies many of these challenges (Henrekson & Wennström, 2022; James & Lewis, 2012). A growing body of research has delved into how public opinion, educational politics, the media, and research reflect and influence perceptions of education (Billingham & Kimelberg, 2016; Lee et al., 2022). However, there is a lack of studies on the interconnectedness of educational discourses using natural data (Lyons, 1991) comprehensively over time. Topic modelling, an established method in historical studies (e.g., Cohen Priva & Austerweil, 2015), can provide valuable opportunities to study discourse formation.

In collaboration with KBLab at the National Library of Sweden (collaboration agreement KB 2024-114), we use exploratory transformer-based topic modelling and sentence-based large language models to analyse extensive data from these discourse areas: the political sphere, the media landscape, and educational sciences. These areas form three corpora represented by 1) parliamentary motions, propositions, and speeches, 2) content from the four largest daily newspapers, and 3) educational research published in scholarly journals. Through KBLab, we are able to obtain close to complete data series from each data source. By dividing the period into 5-year intervals, we investigate relationships within and across these areas to reveal when and where changes are initiated, adopted, and spread to other discourse areas.

Research questions:

1. What changes in the perceptions of knowledge can be found in the last 60 years of political, media, and research discourses?
2. How do changes in these discourses interrelate?

This study contributes to educational research and the social sciences in several ways. Firstly, it provides a robust empirical foundation for exploring the interrelationships among political, public, and scholarly discourses. Secondly, it offers theoretical insights into discourse formation and the processes by which semantic shifts are initiated, adopted, and disseminated across different discourse domains. Thirdly, by fine-tuning (L)LMs, we make a methodological contribution that enables the comprehensive and comparative analysis of extensive, previously inaccessible natural data over time. Additionally, all model codes will be made open-access and available to anyone interested in discourse and policy formation.

This project integrates two methodological traditions: linguistic analysis and statistical regression methods, specifically transformer-based (L)LMs and Granger causality. Combining these methods is crucial as we aim to capture elusive prevailing discourses defined not by a fixed set of lexical items but by 'family-resemblances' of overlapping, non-unique similarities (Wittgenstein, 2001). Simultaneously, we seek to investigate the temporal interrelatedness of these discourses within a structured regression framework. The linguistic analysis will begin with exploratory transformer-based topic modelling for 5-year time intervals within each corpus. We will employ BERTopic, which identifies latent topics by combining a transformer model with traditional information retrieval techniques and density-based clustering (Grootendorst, 2022). In the subsequent analysis, we will use dimensional-reduced results from the topic models to extract, identify, and validate a smaller number of texts with conspicuous loadings for discernible educational topics indicative of potentially prevailing educational discourses. This dataset will serve as training material for fine-tuning a Large Language Model, (L)LM, classifier, allowing the model to 'learn' to classify texts into different discourses. The fine-tuned classifier (L)LM will then perform classification inference on the entire dataset. This classifier will be sentence-based (rather than word-embedded), enabling it to better capture linguistic family resemblances over longer (con-)texts (Reimers & Gurevych, 2019). In the final phase of the study, we will apply Granger causality to analyse the interrelationships between the discourse areas over time (Shojaie & Fox, 2022). Granger causality (Granger, 1980), has been used in various fields to make predictions based on historical data. It indicates a predictive rather than a traditional causal relationship, as previously utilised in scientific analyses of societal discourse shifts (Börner et al., 2018). Through this approach, we aim to investigate whether shifts in one discourse area can predict changes in another, providing insights into the continued development of other domains. For instance, we will examine whether changes to perceptions of knowledge first take hold within research, then influence political debate, eventually appearing in the media, or if alternative patterns exist.

To validate and assess the feasibility of our research, we conducted a pilot study using vector models to compare word embeddings of lexical units related to knowledge perspectives over time in Sweden’s four largest daily newspapers (Dagens Nyheter, Svenska Dagbladet, Aftonbladet, and Expressen). Vector models evaluate the statistical relationships between word pairs on a scale from 0 to 1. Table 1 presents the results of a few examples of word pairs studied during two five-year periods, 1974-1979 and 2004-2009. The results indicate that the relationships are stable over time for some word pairs, such as "content" and "goal," and "teaching" and "learning." For other pairs, the connections have strengthened, indicating closer relationships between the words. Notably, the relationship between "grade" and "test" is significantly stronger in the later period, as is the relationship between "curriculum" and "learning outcomes." These initial models and preliminary results suggest a shift in educational discourse toward more measurable aspects of knowledge. Table 1. Pilot study vector model results (examples of word pairs) Word Pair 1974–1979 2004–2009 Change Sign. Content – Goal 0.37 0.35 -0.02 Teaching – Learning 0.62 0.65 0.03 Formation – Education 0.89 0.85 -0.04 Individual – Group 0.48 0.43 -0.05 Teacher – Pupil 0.75 0.81 0.06 Knowledge – Ability 0.60 0.53 -0.07 Curriculum – Learning Outcomes 0.40 0.55 0.15 * Grade – Test 0.32 0.67 0.35 * The pilot study results support the validity and feasibility of our overarching project, which aims to study educational discourses and discern changes in semantic relationships between words within corpora. At the same time, these results highlight the necessity of employing more advanced linguistic analytical tools and robust transformer-based (L)LMs to comprehensively grasp and understand prevailing educational discourses, which this project aims to achieve.

Billingham, C., & Kimelberg, S. (2016). Opinion polling and the measurement of Americans’ attitudes regarding education. Journal of Education Policy, 31(5), 526–548. Börner, K., Scrivner, O., Gallant, M., Ma, S., Liu, X., Chewning, K., Wu, L., & Evans, J. A. (2018). Skill discrepancies between research, education, and jobs reveal the critical need to supply soft skills for the data economy. Proceedings of the National Academy of Sciences of the United States of America, 115(50), 12630–12637. Cohen Priva, U., & Austerweil, J. L. (2015). Analyzing the history of Cognition using Topic Models. Cognition, 135, 4–9. Greene, J. A., & Yu, S. B. (2016). Educating Critical Thinkers: The Role of Epistemic Cognition. Policy Insights from the Behavioral and Brain Sciences, 3(1), 45–53. Grootendorst, M. (2022). BERTopic: Neural topic modeling with a class-based TF-IDF procedure (arXiv:2203.05794). Henrekson, M., & Wennström, J. (2022). Dumbing Down: The Crisis of Quality and Equity in a Once-Great School System—and How to Reverse the Trend. Springer International Publishing. James, M., & Lewis, J. (2012). Assessment in Harmony with our Understanding of Learning: Problems and Possibilities. In Assessment and Learning (2nd ed., pp. 187–205). SAGE Publications Ltd. Lee, J., Lee, J., & Lawton, J. (2022). Cognitive mechanisms for the formation of public perception about national testing: A case of NAPLAN in Australia. Educational Assessment Evaluation and Accountability, 34(4), 427–457. Lindensjö, B., & Lundgren, U. P. (2000). Utbildningsreformer och politisk styrning (2nd ed.). Liber. Lyons, J. (1991). Natural Language and Universal Grammar: Essays in Linguistic Theory (Vol. 1). Cambridge University Press. Mason, L., Boscolo, P., Tornatora, M. C., & Ronconi, L. (2013). Besides knowledge: A cross-sectional study on the relations between epistemic beliefs, achievement goals, self-beliefs, and achievement in science. Instructional Science, 41(1), 49–79. Muis, K. R., & Foy, M. J. (2010). The effects of teachers’ beliefs on elementary students’ beliefs, motivation, and achievement in mathematics. In Personal Epistemology in the Classroom: Theory, Research, and Implications for Practice (pp. 435–469). Cambridge University Press. Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (arXiv:1908.10084). Schön, L. (2012). An Economic History of Modern Sweden (1st ed.). Routledge. Shojaie, A., & Fox, E. B. (2022). Granger Causality: A Review and Recent Advances. Annual Review of Statistics and Its Application, 9(1), 289–319. Wittgenstein, L. (2001). Philosophical Investigations. Blackwell Publishing.