Maximum Entropy Based Semantic Role Labeling

The semantic role labeling (SRL) refers to finding the semantic relation (e.g. Agent, Patient, etc.) between a predicate and syntactic constituents in the sentences. Especially, with the argument information of the predicate, we can derive the predicateargument structures, which are useful for the applications such as automatic information extraction. As previous work on the SRL, there have been many machine learning approaches. (Gildea and Jurafsky, 2002; Pradhan et al., 2003; Lim et al., 2004). In this paper, we present a two-phase SRL method based on a maximum entropy (ME) model. We first identify parse constituents that represent valid semantic arguments of a given predicate, and then assign appropriate semantic roles to the the identified parse constituents. In the two-phase SRL method, the performance of the argument identification phase is very important, because the argument classification is performed on the region identified at the identification phase. In this study, in order to improve the performance of identification, we try to incorporate clause boundary restriction and tree distance restriction into pre-processing of the identification phase. Since features for identifying arguments are different from features for classifying a role, we need to determine different feature sets appropriate for the tasks. We determine final feature sets for each phase with experiments. We participate in the closed challenge of the CoNLL-2005 shared task and report results on both development and test sets. A detailed description of the task, data and related work can be found in Carreras and Màrquez (2005). 2 System Description