Transition-based Korean Dependency Analysis System Using Semantic Abstraction

ChungSeon Jeong, JoonChoul Shin, JuSang Lee, CheolYoung Ock

http://doi.org/10.5626/JOK.2019.46.11.1174

The existing learning-based dependency studies used as a learning features by combining the lemma and the part-of-speech tag. The part-of-speech tag is suitable for use as a feature due to its high recall, but there is a limit to increase the accuracy of analysis of dependency by using only the part-of-speech tag. In case of lemma, when the lemma is recalled, it shows high dependency accuracy, but it shows low recall compared to the part-of-speech tag. In this paper, we propose a transition-based dependency analysis method that uses abstractions of nouns as a feature by using lexical semantic network (UWordMap) in order to increase the recall rate of lemma. When the semantic abstraction of lemmas is used as a feature, the accuracy of dependency analysis is increased by up to 7.55% compared to the case of using only the lemma. In case of using word(eojeol), morphological and syllable unit features including semantic abstraction features, 90.75% dependence analysis accuracy was shown. With the learning speed of 562 sentences per second and the speed of 631 dependency analysis per second, the proposed method can be used practically.