Unsupervised Korean Word Sense Disambiguation using CoreNet

Unsupervised Korean Word Sense Disambiguation using CoreNet Kijong Han, Sangha Nam, Jiseong Kim, Younggyun Hahm, and Key-Sun Choi Semantic Web Research Center, School of Computing, KAIST 291 Daehak-ro, Yuseong-gu, Daejeon, South Korea {han0ah, nam.sangha, jiseong, hahmyg, kschoi}@kaist.ac.kr Abstract In this study, we investigated unsupervised learning based Korean word sense disambiguation (WSD) using CoreNet, a Korean lexical semantic network. To facilitate the application of WSD to practical natural language processing problems, a reasonable method is required to distinguish between sense candidates. We therefore performed coarse-grained Korean WSD studies while utilizing the hierarchical semantic categories of CoreNet to distinguish between sense candidates. In our unsupervised approach, we applied a knowledge-based model that incorporated a Markov random field and dependency parsing to the Korean language in addition to utilizing the semantic categories of CoreNet. Our experimental results demonstrate that the developed CoreNet based coarse-grained WSD technique exhibited an 80.9% accuracy on the datasets we constructed, and was proven to be effective for practical applications. Keywords: Word Sense Disambiguation, CoreNet, Markov Random Field 1. Introduction Words that have the same form can have different meanings. Word Sense Disambiguation(WSD) is to select the correct meaning of a word in context. It is an important problem that can be utilized in many problems of natural language processing such as machine translation and information extraction (Chaplot et al., 2015). In English, studies are typically conducted on the basis of the senses listed in the Princeton WordNet(PWN) as candidates for the meaning of words (Navigli et al., 2007; Chaplot et al., 2015). In this study, we resolve the ambiguity of words based on the senses listed in CoreNet (Choi et al, 2004), a Korean lexical semantic network. Both PWN and CoreNet are fine-grained resources, so it is difficult for even human annotators to correctly identify the senses of words. In order for WSD to become an enabling technique for end-to-end applications, it requires the ability to make reasonable sense distinctions (Navigli et al., 2007). In English, coarse-grained WSD is performed by semi- automatically clustering PWN senses (Navigli et al., 2007). However, Korean dictionaries also list homograph and polyseme numbers for each sense. Homographs are words that have the same form but completely different meanings, and polysemes are words that have the same broad meaning but different etymological meanings. For this reason, in Korean language studies, coarse-grained WSD is typically conducted at the homograph level (Shin and Ock, 2016). WSD methods can generally be divided into supervised and unsupervised learning methods. Supervised learning methods learn sense-tagged corpora and show relatively high levels of performance. However, it is expensive and time-consuming to construct such corpora. The results of one Korean language study showed a 96.5% accuracy based on learning a corpus consisting of 10 million sense tagged words (Shin and Ock, 2016). In contrast, unsupervised learning methods are relatively low performance but do not require training data. For example, knowledge-based methods that employ lexical semantic networks have been successful (Agirre et al., 2014; Chaplot et al., 2015) because they are …

Content truncated to 500 words