Towards Fully Lexicalized Dependency Parsing for Korean

Abstract

We propose a Korean dependency parsing system that can learn the relationships between Korean words from the Treebank corpus and a large raw corpus. We first refine the training dataset to better represent the relationship using a different POS tagging granularity type. We also introduce lexical information and propose an almost fully lexicalized probabilistic model with case frames automatically extracted from a very large raw corpus. We evaluate and compare systems with and without POS granularity refinement and case frames. The proposed lexicalized method outperforms not only the baseline systems but also a state-of-the-art supervised dependency parser.