Exploring Word Embedding in Modeling Risk Perception

Risk perception models attempt to capture how individuals or groups perceive and evaluate risks, focusing on subjective assessments rather than quantitative probabilities of occurrence. The aim is to identify influencing factors that may affect risk perception with the greater aim of better understanding (cognitive) decision-making processes in risk scenarios. To operationalise these insights, we apply word embedding techniques to quantify and assess risk perception information embedded in textual data. This approach, based on large corpora and machine learning models, allows predicting risk perception from the semantic content of language(s). This paper complements initial approaches of using word embedding vector space to forecast risk perceptions scores. We deepened that understanding by examining how the use of similar words to given risk terms in a vector space, or context words can provide additional information regarding their semantic use. Finally, we applied arithmetic operations to incorporate cultural and geographical contexts into risk perception. The results show that adding more words to create context reduces performance of the models, while using arithmetic operations can provide better forecasts of risk perceptions, and that they may also be used to further explore cultural or geographic variations in risk perceptions.