Conceptual structure and the growth of scientific knowledge

Abstract

How does scientific knowledge grow? This question has occupied a central place in the philosophy of science, stimulating heated debates, but yielding no clear consensus. Many explanations can be understood in terms of whether and how they view the expansion of knowledge as proceeding through the accretion of scientific concepts into larger conceptual structures. Here, we examine these views empirically, performing a large-scale analysis of the physical and social sciences, spanning five decades. Using natural language processing techniques, we create semantic networks of concepts, wherein noun phrases become linked when used in the same paper abstract. For both the physical and social sciences, we observe increasingly rigid conceptual cores (i.e., densely connected sets of highly central nodes) accompanied by the proliferation of periphery concepts (i.e., sparsely connected nodes that are highly connected to the core). Subsequently, we examine the relationship between conceptual structure and the growth of scientific knowledge, finding that scientific works are more innovative in fields with cores that have higher conceptual churn and with larger cores. Furthermore, scientific consensus is associated with reduced conceptual churn and fewer conceptual cores. Overall, our findings suggest that while the organization of scientific concepts is important for the growth of knowledge, the mechanisms vary across time.

Figures (13)

• Figure 1: Schematic illustration of core/periphery measures over time for two fields. (1A) illustrates different ways in which core/periphery structures can unfold across time. The highly connected nodes at the center of each network represent the core, whereas the relatively-less connected nodes toward the edge of the networks represent the periphery. The color of the nodes indicates the nodes' status from the prior year---pink indicates that the node was in the core, purple indicates that the node was in the periphery, and blue indicates that the node was added to the network at time t. At t=1, Field A and Field B are identical. As time progresses, each field differs in terms of our core/periphery measures. Field A maintains a single core across time, while additional cores emerge in Field B (i.e., there are two cores at t=2 and four cores at t=3). The fields also differ in terms of the churn in core nodes. Field A has low churn; the pink-core nodes remain in the core and a small number of purple-periphery nodes and blue-new nodes are added to the core. In contrast, the core nodes in Field B change over time; that is, purple nodes—which were in the periphery during the prior year—and blue nodes—which are new to the network—enter the core. Finally, the fields differ in terms of the relative size of the core. Field A has a relatively smaller core, driven by growth in the periphery. Field B has a relatively larger core, with similar numbers of core and periphery nodes. (1B) plots how the fields change in terms of the core/periphery measures across time.
• Figure 2: Conceptual structures over time. The plots show changes in the core/periphery organization averaged across subfields. For both the social sciences (top) and physical sciences (bottom), the plots show a decrease in the amount of churn of core concepts (a,d) and a decrease in the relative number of core nodes (b,e). However, the social and physical sciences differ in terms of the number of cores over time. We see a slight decrease in the number of cores for the social sciences (c). Alternatively, we see an overall increase in the number of cores for the physical sciences (f). The inset plots (found in plots c and f) show that concepts become increasingly concentrated in a few number of core/periphery structures across time.
• Figure 3: Schematic illustration of core/periphery measures.(1A) Comparing the conceptual structure of Sociology in 1995 and 2010 reveals that the relative size of the core decreases over time. Specifically, there are more periphery nodes in 2010 (visually indicated as smaller in size) causing the relative size of the core to shrink. (1B) Comparing the structure of Archaeology in 1995 and 2010 reveals a decrease in the churn of core concepts. Core nodes in the 1995 network include many new concepts (blue) and prior periphery concepts (purple) in addition to prior core concepts (pink). In contrast, the core in the 2010 network is composed primarily of nodes that were previously in the core (pink). (1C) The conceptual structure of Women's Studies 2010 provides an example of a multi-core network, wherein the majority of concepts are concentrated around two large core/periphery structures, one corresponding to more humanistic approaches (bottom left), and the other more biomedical ones (top right). Note: Due to the computational burden of generating large-scale network diagrams, we could not incorporate all the concepts present in our analytical data into this visual representation.
• Figure 4: Concepts extracted from the text of an abstract. This figure shows an example abstract from the APS data; the highlighted text indicates single-word and multi-word noun phrases identified as concepts using our extraction algorithm.
• Figure 5: Mobility of concepts to/from the core/periphery over time. To generate this figure, we calculated the proportion of core nodes at year $t$ that either stayed in the core, transitioned to the periphery, or exited the network at year $t+1$. Our results show that core concepts are more likely to stay in the core at year $t+1$ for both the social sciences (top, left) and physical sciences (bottom, left). Further, we find that if core concepts exit the core at year $t+1$, they are more likely to enter the periphery than exit the network. Next, we calculated the proportion of periphery nodes at year $t$ that either stayed in the periphery, transitioned to the core, or exited the network at year $t+1$. Our results show that periphery concepts at year $t+1$ are more likely to stay in the periphery as time progresses for the social sciences (top, right) and for the physical sciences (bottom, right). We also observe that if periphery concepts exit the periphery, they are more likely to enter the core than exit the network.