Semantic Dependency Parsing of NomBank and PropBank: An Efficient Integrated Approach via a Large-scale Feature Selection

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 30–39, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Semantic Dependency Parsing of NomBank and PropBank: An Efficient Integrated Approach via a Large-scale Feature Selection ∗ Hai Zhao(赵 赵 赵海 海 海)†, Wenliang Chen(陈 陈 陈文 文 文亮 亮 亮)∗, Chunyu Kit†(揭 揭 揭春 春 春雨 雨 雨) †Department of Chinese, Translation and Linguistics City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong, China ∗Language Infrastructure Group, MASTAR Project National Institute of Information and Communications Technology 3-5 Hikari-dai, Seika-cho, Soraku-gun, Kyoto, Japan, 619-0289 haizhao@cityu.edu.hk, chenwl@nict.go.jp Abstract We present an integrated dependency- based semantic role labeling system for English from both NomBank and Prop- Bank. By introducing assistant argument labels and considering much more fea- ture templates, two optimal feature tem- plate sets are obtained through an effec- tive feature selection procedure and help construct a high performance single SRL system. From the evaluations on the date set of CoNLL-2008 shared task, the per- formance of our system is quite close to the state of the art. As to our knowl- edge, this is the first integrated SRL sys- tem that achieves a competitive perfor- mance against previous pipeline systems. 1 Introduction We investigate the possibility to construct an effec- tive integrated system for dependency-based se- mantic role labeling (SRL) task. This means in this work that a single system handles all these sub-tasks, predicate identification/disambiguation and argument identification/classification, regard- less of whether the predicate is verbal or nominal. Traditionally, a SRL task, either dependency or constituent based, is implemented as two sub- tasks, namely, argument identification and clas- sification. If the predicate is unknown, then a predicate identification or disambiguation subtask should be additionally considered. A pipeline framework is usually adopted to handle all these sub-tasks. The reason to divide the whole task ∗This study is partially supported by CERG grant 9040861 (CityU 1318/03H), CityU Strategic Research Grant 7002037. into multiple stages is two-fold, one is each sub- task asks for its favorable features, the other is at the consideration of computational efficiency. Generally speaking, a joint system is slower than a pipeline system in training. (Xue and Palmer, 2004) fount out that different features suited for different sub-tasks of SRL, i.e. argument identifi- cation and classification. The results from CoNLL shared tasks in 2005 and 2008 (Carreras and Mar- quez, 2005; Koomen et al., 2005; Surdeanu et al., 2008; Johansson and Nugues, 2008), further show that SRL pipeline may be one of the standard to achieve a state-of-the-art performance in practice. In the recent years, most works on SRL, includ- ing two CoNLL shared task in 2004 and 2005, focus on verbal predicates with the availability of PropBank (Palmer et al., 2005). As a com- plement to PropBank, NomBank (Meyers et al., 2004) annotates nominal predicates and their cor- responding semantic roles using similar semantic framework as PropBank. Though SRL for nomi- nal predicates offers more challenge, it draws rel- atively little attention (Jiang …

Content truncated to 500 words