Semi-automatic Korean FrameNet Annotation over KAIST Treebank

Semi-automatic Korean FrameNet Annotation over KAIST Treebank Younggyun Hahm, Sunggoo Kwon, Jiseong Kim, Key-Sun Choi Semantic Web Research Center, School of Computing, KAIST 291 Daehak-ro, Yuseong-gu, Deajeon, South Korea {hahmyg, fanafa, jiseong, kschoi}@kaist.ac.kr Abstract This paper describes a project for constructing FrameNet annotations in Korean over the KAIST treebank corpus to scale up the Korean FrameNet resource. Annotating FrameNet over raw sentences is an expensive and complex task, because of which we have designed this project using a semi-automatic annotation approach. This paper describes the approach and its expected results. As a first step, we built a lexical database of the Korean FrameNet, and used it to learn the model for automatic annotation. Its current scope, status, and limitations are discussed in this paper. Keywords: FrameNet, Semantic Role Labeling, Corpus Annotation 1. Introduction FrameNet is a large lexical database that has rich annotations to represent the meanings of text using semantic frames (Baker et al., 1998; Fillmore et al., 2003). FrameNet has been considered a useful resource for various applications such as question answering systems (Shen and Lapata, 2007, Hahm et al., 2016), information extraction (Surdeanu et al., 2003), and dialog systems (Chen et al., 2013). Lately, researchers have shown increasing interest in multilingual FrameNet (Borin et al., 2010; You and Lui, 2005; Meurs et al., 2008; Subirats and Petruck, 2003; Burchardt et al., 2006). A Korean FrameNet project built the Korean FrameNet resource by translating English FrameNet annotations and Japanese FrameNet into its equivalent in Korean (Park et al., 2014; Kim et al., 2016). One of the purpose of the FrameNet annotation is to build a frame-semantic parser to understand the meaning of a text. Some studies have built frame-semantic parser for English by using full-text annotation and partially annotated exemplar sentences to train their models (Das et al., 2010; Swayamdipta et al., 2017; Yang and Mitchell, 2017). For example, the state-of-the-art frame-semantic parser uses nearly 139k exemplar sentences for training data and it generally introduces a 3–4 F1 gain for parsing (Yang and Mitchell., 2017). In comparison, the Korean FrameNet has full-text annotations for 5,025 sentences and 8,200 lexical units (LUs). In this paper, we report an ongoing project to construct Korean FrameNet annotations to scale up the amount of annotations. A task to annotate the frame-semantics manually over raw sentences has been formalized by Ruppenhofer et al. (2006). It is an expensive and complex task, because the annotators would need to choose proper frame-semantics for each target word, and its corresponding frame elements for the arguments (1,221 frame-semantics are defined in FrameNet 1.7). Therefore, we have used the existing Korean resources to bootstrap the annotation task. A target corpus for the FrameNet annotation task is the KAIST treebank (Choi et al., 1994). It includes 31,086 sentences with morphological analysis, part-of-speech (POS) tagging, and dependency parsing. It is used in the Korean Universal Dependency treebank (Choi, 2013). This paper briefly introduces the FrameNet annotation over the KAIST treebank project, and presents its scope in Section 2. The related …

Content truncated to 500 words