Automatic Semantic Role Labelling (SRL) in Swedish

Abstract

In this paper, using deep learning networks, the first end-to-end semantic role labelling model (SRL) has been developed for Swedish texts. This Swedish SRL model can, with a given Swedish sentence, perform trigger identification, frame classification and argument extraction tasks automatically in a series. Using the semantic annotation examples from the Swedish FrameNet corpus (SweFN), we experiment with two transfer learning approaches. We show that the Swedish SRL based on a pre-trained English SRL can speed up the training, meanwhile the multilingual model (mT5) based model has better performance when the possible frames are unknown. Through extensive empirical analysis of the model performance, we point out the major factors that can further improve the results.