SASHIMI and new frontiers in the study of socio-semantic networks with mixed-methods on the Cortext Platform

Résumé En

Since 2008, the Cortext Platform contributes expertise, infrastructure and computing power for the analysis of "socio-semantic networks", benefiting a global community engaged primarily in original research in the social sciences and humanities, but also assisting literature reviews in a host of others fields, as well as policy and business applications. In 2022, at least 60 peer-reviewed academic publications made direct use of our services, adding to a total of over 300. This presentation will focus on SASHIMI (Hannud Abdo, 2021), a network based, mixed-methods approach recently developed in addition to our earlier Network Mapping methods (Cointet 2012, Cointet 2017), available as both a suite of no-code methods in the free-to-use Cortext Manager cloud service, and a free-and-open-source software library. We will present SASHIMI through some examples of socio-semantic analyses: (a) from the field of Transition Studies, an inquiry into the variety of disciplinary manifestations throughout the social sciences of the "research problem of destabilisation of socio-technical systems", that seeks to inform current destabilisation/discontinuation/phase-out studies with a wider understanding of the problem. (b) from the field of Science and Technology Studies, an analysis of policy documents pertaining to the regulation of artificial intelligence, identifying the interplay between major actors associated with different themes, sectors and perspectives (solutionism, contestation, regulation) on the issue; (c) still in STS, an analysis of social media interactions concerning environmental controversies, focusing on the debate around pesticides in France. SASHIMI is based on domain-topic models, an application of network clustering that synthesizes document clustering (or clustering of any kind of hypernode) and topic modeling. It is also based on a suite of human interfaces — block maps, network maps, and hyperlinked tables — that afford interactive exploration and visualization of the different types of clusters, and their relationships, at discrete levels of granularity ranging from the entire corpus to the individual document, from the entire vocabulary to the individual word. The clustering aspect is based on modern community detection methods, namely the Nested Stochastic Block Model (Peixoto, 2015), while introducing a twist to allow further clustering of dimensions attributed to hypernodes (documents), such as people, time, venue or other categorical metadata, that did not participate in the initial clustering — excluded, for example, in order to produce "semantic" document clusters based exclusively on textual contents. To this particular procedure we give the name "chaining". In the context of the three aforementioned examples, we'll explain a set of concepts and practices, emerging from our usage, to productively co-construct meaning between the representations afforded by the models and interfaces, and the goals, inputs and choices of a researcher with field and experiential knowledge. Particularly, how to interpret the clusters and the specificity and commonality scores of inter-cluster relationships employed in the maps, how to build sequences of corpus delimitation and dimension chaining operations and interpret them, and finally how to construct coherent domain groups we call "constellations", and identify attribute flows in their cores and frontiers.