ILK2: Semantic Role Labeling of Catalan and Spanish using TiMBL

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 183–186, Prague, June 2007. c⃝2007 Association for Computational Linguistics ILK2: Semantic Role Labelling for Catalan and Spanish using TiMBL Roser Morante, Bertjan Busser ILK, Dept. of Language and Information Sciences Tilburg University, P.O.Box 90153 NL-5000 LE Tilburg, The Netherlands {R.Morante,G.J.Busser}@uvt.nl Abstract In this paper we present a semantic role la- beling system submitted to the task Multi- level Semantic Annotation of Catalan and Spanish in the context of SemEval–2007. The core of the system is a memory–based classifier that makes use of full syntactic in- formation. Building on standard features, we train two classifiers to predict separately the semantic class of the verb and the seman- tic roles. 1 Introduction Semantic role labelling (SRL) has been addressed in the CoNLL–2004 and CoNLL–2005 Shared Tasks (Carreras and M`arquez, 2004; Carreras and M`arquez, 2005) for English. In the task Multilevel Semantic Annotation of Catalan and Spanish of the SemEval competition 2007, the target are two differ- ent languages. The general SRL task consists of two tasks: prediction of semantic roles (SR) and predic- tion of the semantic class of the verb (SC). The data provided in the task (M`arquez et al., 2007) are sentences annotated with lemma, POS tags, syntactic information, semantic roles, and the semantic classes of the verb. A training corpus for Catalan (ca.3LB) and another for Spanish (sp.3LB) are provided. Although the setting is similar to the CoNLL–Shared Task 2005, three relevant differ- ences are that the corpora are significantly smaller, that the syntactic information is based on a manu- ally corrected treebank, which contains also syntac- tic functions (i.e. direct object, indirect object, etc.), and that the set of semantic roles is larger, especially for core arguments. Our goal is to check whether simple individual systems could produce competitive results in both subtasks, and whether they would be robust enough when applied to two languages and to the held–out test sets provided. 2 System description We approach the SRL task as two classification problems: prediction of SR and prediction of SC. We hypothesize that the two problems can be solved in the same way for both languages. We build two very similar systems that differ only in some of the features used, as we explain below. The task is solved in three phases: 1) A pre– processing phase that is very similar to the sequen- tialization in (M`arquez et al., 2005). We call it focus selection. It consists of identifying the potential can- didates to be assigned a semantic role or a semantic verb class. 2) The classification. 3) Some limited postprocessing. 2.1 Focus selection The system starts by finding the target verb (which is marked in the corpus as such). Then, it finds the complete form of the verb (that in the corpus is tagged as verb group, infinitive, gerund, etc.) and the clause boundaries in order to look for the siblings of the verb that are under the same clause. Our as- sumption is that …

Content truncated to 500 words