Daejeon, July 27–31
While various methods exist to model change in language use, interpretability remains a difficult challenge (Periti et al., 2024; Karjus, 2025). This hinders progress in qualitative evaluations and model adaptations by domain experts (Beck, 2024). As a case study, we analyze two competing theories during the Chemical Revolution: phlogiston vs. oxygen (1750-1800). Considering these theories as prototypical concepts (Geeraerts, 1997), our central question is whether the network structure reflects the spread and competition of readings in changing scientific concepts.
Methods. Based on the Royal Society Corpus (RSC; Kermes et al., 2016; Fischer et al., 2020) this approach consisted of four stages: first, create a representative corpus by filtering publications given specified decades and oxygen terminology (Chang, 2011; Degaetano-Ortlieb et al., 2019; Teich et al., 2020; Bizzoni et al. 2021). Next, represent concepts and their diachronic readings using topic modeling (Blei et al., 2003; Sievert et al., 2014). Then, form concept networks by linking documents based on the Jensen-Shannon distance (Shannon, 1948) among topics and optimize connectivity by minimizing the percolation threshold (Radicchi, 2015); this is the tipping point where nodes in a graph reach optimal connectivity with a minimum number of edges, where a high threshold indicates maintaining connectivity requires maximal resources (see Figure 3). Finally, reveal hidden groups by applying clustering (Lukasová, 1979) and community detection (Blondel et al., 2008) algorithms.
Discussion. Figure 1 shows onomasiological change, where we consider topic labels as prototype concepts with central (top-10) and peripheric (top-50) readings. Notably, while the prototype concept “air” takes “acid” as a central reading in the 1780s, this relationship flips by the 1800s when “acid” becomes prototypical with “oxygen” acting as a peripheric reading. Note that the cumulative strategy, containing the aggregation of papers until a target decade, shows topic clusters with recurring labels (e.g., “plant”, “air”), while the non-cumulative, containing the papers per decade only, offered fine-grained representations of the themes discussed.
These observations are consistent with the conceptual structure reported in the literature where the term “acid” played a pivotal role for the conceptual change (Thagard, 1990; Chang, 2011): while phlogistonists like Priestley underplayed acidity as a secondary effect of phlogiston release (e.g., experiments burning sulfur produced a “gaseous calx” that was perceived as acidic since it lacked phlogiston), Lavoisier considered acidity as a core explanatory principle to reframe calxes as oxides and oxygen as an enabler of acid formation (e.g., sulfur is a combustible that when in contact with oxygen forms sulfur dioxide, which in turn dissolves in water to produce sulfurous acid).
To compare the magnitude of diachronic change in both sampling strategies, we measured the topics “diversity of ideas” using entropy (Griffiths et al., 2004). Figure 2 presents different trends: while the cumulative strategy shows rising entropy across oxygen-related topics, the non-cumulative declines then rises post-1774 oxygen discovery. This decline of diversity pre-1780 signals the resistance of phlogistonists to decenter “air” in their findings (e.g., Priestley’s “dephlogisticated air”), while the rise of diversity post-1780 originates from Lavoisier coining the term “oxygen”. These observations support Thagard (1990) on why Lavoisier made the conceptual shift instead of Priestley. Future studies will address Lavoisier’s articles in French to clarify this.
Considering the onomasiological change (Figure 1) and rising entropy (Figure 2), we built graphs based on the non-cumulative strategy producing 6 in total (Figure 3). To analyze network stability, we considered 5 parameters (Figure 4), from where we draw these observations:
In conclusion, after the discovery of oxygen: 1) a topic cluster shifts from “air” to “acid,” showing onomasiological change given the movement of core/peripheric readings accompanied by higher entropy, and 2) network communities and modularity decline, while percolation and edge density rise, indicating higher connectivity effort. These results suggest changing concepts form high entropy clusters that increase connectivity effort in the knowledge network. Arguably, this framework generates empirical evidence for theories in philosophy of science, such as the decline in entropy pre-1780 being a signal of resistance toward counter-establishment ideas (Fleck, 1981). Future work aims to investigate this direction via directed graphs, rare-term augmentation, and estimating phylogenetic structures among concepts/authors.
Figure 1. Topic clusters: 1780 vs. 1800. Cumulative sampling shows stable labels over time, contrary to the non-cumulative with more fine-grained representations.
Figure 2. Topic entropy. While the cumulative strategy shows rising entropy across oxygen-related topics, the non-cumulative declines then rises post-1774 oxygen discovery.
Figure 3. Temporal graphs. We found that the number of communities—here illustrated with colors—went on decline: starting at 5 (1750s), then 4 (1770s), and ending at 3 (1780s and 1800s).
Figure 4. Network metrics. Nodes size, edge density (where y-axis=1x103), communities count, modularity and percolation.