Enlisting supervised machine learning in mapping scientific uncertainty expressed in food risk analysis
Résumé
Recently, both sociology of science and policy research have shown increased interest in scientific uncertainty. To contribute to these debates and create an empirical measure of scientific uncertainty, we inductively devised two systems of classification or ontologies to describe scientific uncertainty in a large corpus of food safety risk assessments with the help of machine learning (ML). We ask three questions: (1) Can we use ML to assist with coding complex documents such as food safety risk assessments on a difficult topic like scientific uncertainty? (2) Can we assess using ML the quality of the ontologies we devised? (3) And, finally, does the quality of our ontologies depend on social factors? We found that ML can do surprisingly well in its simplest form identifying complex meanings, and it does not benefit from adding certain types of complexity to the analysis. Our ML experiments show that in one ontology which is a simple typology, against expectations, semantic opposites attract each other and support the taxonomic structure of the other. And finally, we found some evidence that institutional factors do influence how well our taxonomy of uncertainty performs, but its ability to capture meaning does not vary greatly across the time, institutional context, and cultures we investigated.