Speaker: Denitsa Saynova from Chalmers University of Technology
Title: Explainability for text models in the context of political science research
Abstract: Can we identify party affiliation from debates in the Swedish Riksdag with supervised learning? What can explainability methods reveal about the distinguishing features of the respective parties? This work aims to examine the advancements within explainability for text models and the alignment with use-cases in political science. Two main gaps are identified. First - the understudied area of class explanations - explanations for properties of groups rather than single input instances. Second - the type of features used for providing explanations. Majority of current methods focus on feature importance, which in text modelling most often corresponds to single words. It is not clear how we can overlay more complex structures as part of the explanations.