UBC-UPC: Sequential SRL Using Selectional Preferences. An approach with Maximum Entropy Markov Models

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 354–357, Prague, June 2007. c⃝2007 Association for Computational Linguistics UBC-UPC: Sequential SRL Using Selectional Preferences. An aproach with Maximum Entropy Markov Models Be˜nat Zapirain, Eneko Agirre IXA NLP Group University of the Basque Country Donostia, Basque Country {benat.zapirain,e.agirre}@ehu.es Llu´ıs M`arquez TALP Research Center Technical University of Catalonia Barcelona, Catalonia lluism@lsi.upc.edu Abstract We present a sequential Semantic Role La- beling system that describes the tagging problem as a Maximum Entropy Markov Model. The system uses full syntactic in- formation to select BIO-tokens from input data, and classifies them sequentially us- ing state-of-the-art features, with the addi- tion of Selectional Preference features. The system presented achieves competitive per- formance in the CoNLL-2005 shared task dataset and it ranks first in the SRL subtask of the Semeval-2007 task 17. 1 Introduction In Semantic Role Labeling (SRL) the goal is to iden- tify word sequences or arguments accompanying the predicate and assign them labels depending on their semantic relation. In this task we disambiguate ar- gument structures in two ways: predicting VerbNet (Kipper et al., 2000) thematic roles and PropBank (Palmer et al., 2005) numbered arguments, as well as adjunct arguments. In this paper we describe our system for the SRL subtask of the Semeval2007 task 17. It is based on the architecture and features of the system named ‘model 2’ of (Surdeanu et al., forthcoming), but it introduces two changes: we use Maximum Entropy for learning instead of AdaBoost and we enlarge the feature set with combined features and other seman- tic features. Traditionally, most of the features used in SRL are extracted from automatically generated syntac- tic and lexical annotations. In this task, we also ex- periment with provided hand labeled semantic infor- mation for each verb occurrence such as the Prop- Bank predicate sense and the Levin class. In addi- tion, we use automatically learnt Selectional Prefer- ences based on WordNet to generate a new kind of semantic based features. We participated in both the “close” and the “open” tracks of Semeval2007 with the same system, mak- ing use, in the second case, of the larger CoNLL- 2005 training set. 2 System Description 2.1 Data Representation In order to make learning and labeling easier, we change the input data representation by navigating through provided syntactic structures and by extract- ing BIO-tokens from each of the propositions to be annotated as shown in (Surdeanu et al., forthcom- ing). These sequential tokens are selected by ex- ploring the sentence spans or regions defined by the clause boundaries, and they are labeled with BIO tags depending on the location of the token: at the beginning, inside, or outside of a verb argument. Af- ter this data pre-processing step, we obtain a more compact and easier to process data representation, making also impossible overlapping and embedded argument predictions. 2.2 Feature Representation Apart from Selectional Preferences (cf. Section 3) and those extracted from provided semantic infor- mation, most of the features we used are borrowed …

Content truncated to 500 words