Feature Selection Using a Semantic Hierarchy for Event Recognition and Type Classification

International Joint Conference on Natural Language Processing, pages 136–144, Nagoya, Japan, 14-18 October 2013. Feature Selection Using a Semantic Hierarchy for Event Recognition and Type Classification Yoonjae Jeong and Sung-Hyon Myaeng Korea Advanced Institute of Science and Technology (KAIST) 291 Daehak-ro (373-1 Guseong-dong), Yuseong-gu, Daejeon 305-701, Republic of Korea {hybris, myaeng}@kaist.ac.kr Abstract Event recognition and event type classifi- cation are among the important areas in text mining. A state-of-the-art approach utilizing deep-level lexical semantics and syntactic dependencies suffers from a limitation of requiring too large feature space. In this paper, we propose a novel feature selection method using a semantic hierarchy of features based on WordNet relations and syntactic dependencies. Compared to the well-known feature se- lection methods, our proposed method reduces the feature space significantly while keeping the same level of effec- tiveness. For noun events, it improves ef- fectiveness as well as efficiency. Moreo- ver, we expect the proposed feature se- lection can be applied to the other types of text classification using hierarchically organized semantic resources such as WordNet. 1 Introduction Feature selection is an important issue in text- based classification because features can be gen- erated in a number of different ways from text. Selecting features affects not only efficiency when the space is big but also classification ef- fectiveness by eliminating noise features (Manning, Raghavan, & Schütze, 2008). In this paper, we propose a new feature selection meth- od that utilizes semantic aspects of word features and discuss its relative merits compared to other well-known feature selection methods. Among many text-based classification prob- lems, this research focuses on event recognition (a kind of binary classification) and type classifi- cation that have been studied extensively to im- prove performance of applications such as auto- matic summarization (Daniel, Radev, & Allison, 2003) and question answering (Pustejovsky, 2002). For event recognition and type classifica- tion, TimeML has served as a representative an- notation scheme of events (Pustejovsky, Castaño, et al., 2003), which are defined as situations that happen or occur and expressed by verbs, nomi- nalizations, adjectives, predicative clauses or prepositional phrases. TimeML defines seven types of events, REPORTING, PERCEPTION, ASPECTUAL, I_ACTION, I_STATE, STATE, and OCCURRENCE (Pustejovsky, Knippen, Littman, & Saurí, 2007), to which a recognized event text is classified for event type classifica- tion. Different approaches to recognize and classify TimeML events have been proposed, ranging from rule-based approaches (Saurí, Knippen, Verhagen, & Pustejovsky, 2005) to supervised machine learning techniques based on lexical semantic classes and morpho-syntactic infor- mation around events (Bethard & Martin, 2006; Boguraev & Ando, 2007; Jeong & Myaeng, 2013; Llorens, Saquete, & Navarro-Colorado, 2010). Jeong & Myaeng (2013) recently showed that using the deeper-level of semantics in- creased the performance. They obtained the best performance in their classification experiments when lexical semantic features using hypernyms at the maximum depth of eight in WordNet were used for the event candidates and the words hav- ing syntactic dependency. While the approach showed a meaningful improvement, it has a problem of generating too many features. Semantic …

Content truncated to 500 words