Improving Semantic Frame Accuracy for Semantic Dependency Parsing

Abstract

This thesis presents an in-depth contrastive error analysis of a set of semantic dependency parsing systems. Based on the empirical results of our analysis we found semantic frame classification to be an interesting case study. As part of this thesis we have made a semantic frame classifier that outperforms previous results. The semantic frame classifier is the result of rigorous experimentation with four set of features: (1) lexical, (2) morphological, (3) syntactic, and (4) semantic. We show that our results outperform previous results. We also show that our classifier can be used to extend and improve the frame semantic classification accuracy of two existing state-of-the-art semantic dependency parsing systems.