Korean Dependency Parsing with Proper Noun Encodin

Abstract

Dependency parsing is the process of analyzing the linguistic relationship of words that make up a sentence. In natural language processing using deep learning, previously registered words are represented in a continuous vector space. However, if proper noun exists in a sentence, it is difficult to represent it in a continuous vector space, and there may be a problem in the dependency parsing. To solve these problems, we propose a dependency parsing method including proper noun in Korean. Before representing words in a continuous vector space, we replaced the proper noun with a special token and learned the environmental features using the multilayer bidirectional LSTM. We distinguished neural networks according to the presence of proper noun, and compared the performance with the Malt parser using the same transition-based method for the entire sentence. We show that the proposed model is 1.7%p better in UAS and 2.1%p better in LAS than Malt parser.