Semantic Role Labeling

Book chapter, 2021

We investigate the feasibility of automatic semantic role labeling (SRL) using Swedish FrameNet (SweFN). In the first part of the chapter, we describe a baseline system using a traditional division into segmentation and labeling steps. These subsystems are implemented as separate machine learning models, and we explore a wide range of syntactic and lexical features for these models. In the second part, we turn to the question of how the frame-to-frame relations defined in FrameNet allow us to use the annotated examples more effectively. The cross-frame generalization methods reduce the number of errors made by the labeling classifier by 27%. For previously unseen frames, the reduction is even more significant: 50%.