Learning Predictive Structures for Semantic Role Labeling of NomBank

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 208–215, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Learning Predictive Structures for Semantic Role Labeling of NomBank Chang Liu and Hwee Tou Ng Department of Computer Science National University of Singapore 3 Science Drive 2, Singapore 117543 {liuchan1, nght}@comp.nus.edu.sg Abstract This paper presents a novel application of Alternating Structure Optimization (ASO) to the task of Semantic Role Labeling (SRL) of noun predicates in NomBank. ASO is a recently proposed linear multi-task learn- ing algorithm, which extracts the common structures of multiple tasks to improve accu- racy, via the use of auxiliary problems. In this paper, we explore a number of different auxiliary problems, and we are able to sig- nificantly improve the accuracy of the Nom- Bank SRL task using this approach. To our knowledge, our proposed approach achieves the highest accuracy published to date on the English NomBank SRL task. 1 Introduction The task of Semantic Role Labeling (SRL) is to identify predicate-argument relationships in natural language texts in a domain-independent fashion. In recent years, the availability of large human-labeled corpora such as PropBank (Palmer et al., 2005) and FrameNet (Baker et al., 1998) has made possible a statistical approach of identifying and classifying the arguments of verbs in natural language texts. A large number of SRL systems have been evalu- ated and compared on the standard data set in the CoNLL shared tasks (Carreras and Marquez, 2004; Carreras and Marquez, 2005), and many systems have performed reasonably well. Compared to the previous CoNLL shared tasks (noun phrase bracket- ing, chunking, clause identification, and named en- tity recognition), SRL represents a significant step towards processing the semantic content of natural language texts. Although verbs are probably the most obvious predicates in a sentence, many nouns are also ca- pable of having complex argument structures, often with much more flexibility than its verb counterpart. For example, compare affect and effect: [subj Auto prices] [arg−ext greatly] [pred affect] [obj the PPI]. [subj Auto prices] have a [arg−ext big] [pred effect] [obj on the PPI]. The [pred effect] [subj of auto prices] [obj on the PPI] is [arg−ext big]. [subj The auto prices’] [pred effect] [obj on the PPI] is [arg−ext big]. The arguments of noun predicates can often be more easily omitted compared to the verb predi- cates: The [pred effect] [subj of auto prices] is [arg−ext big]. The [pred effect] [obj on the PPI] is [arg−ext big]. The [pred effect] is [arg−ext big]. With the recent release of NomBank (Meyers et al., 2004), it becomes possible to apply machine learning techniques to the task. So far we are aware of only one English NomBank-based SRL system (Jiang and Ng, 2006), which uses the maximum entropy classifier, although similar efforts are re- ported on the Chinese NomBank by (Xue, 2006) 208 and on FrameNet by (Pradhan et al., 2004) us- ing a small set of hand-selected nominalizations. Noun predicates also appear in FrameNet semantic role …

Content truncated to 500 words