Improving Semantic Role Classification with Selectional Preferences

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 373–376, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Improving Semantic Role Classification with Selectional Preferences Be˜nat Zapirain, Eneko Agirre IXA NLP Group Basque Country Univ. {benat.zapirain,e.agirre}@ehu.es Llu´ıs M`arquez TALP Research Center Technical Univ. of Catalonia lluism@lsi.upc.edu Mihai Surdeanu Stanford NLP Group Stanford Univ. mihais@stanford.edu Abstract This work incorporates Selectional Prefer- ences (SP) into a Semantic Role (SR) Clas- sification system. We learn separate selec- tional preferences for noun phrases and prepo- sitional phrases and we integrate them in a state-of-the-art SR classification system both in the form of features and individual class predictors. We show that the inclusion of the refined SPs yields statistically significant im- provements on both in domain and out of do- main data (14.07% and 11.67% error reduc- tion, respectively). The key factor for success is the combination of several SP methods with the original classification model using meta- classification. 1 Introduction Semantic Role Labeling (SRL) is the process of extracting simple event structures, i.e., “who” did “what” to “whom”, “when” and “where”. Current systems usually perform SRL in two pipelined steps: argument identification and argument classification. While identification is mostly syntactic, classifica- tion requires semantic knowledge to be taken into account. Semantic information is usually captured through lexicalized features on the predicate and the head–word of the argument to be classified. Since lexical features tend to be sparse, SRL systems are prone to overfit the training data and generalize poorly to new corpora. Indeed, the SRL evaluation exercises at CoNLL- 2004 and 2005 (Carreras and M`arquez, 2005) ob- served that all systems showed a significant perfor- mance degradation (∼10 F1 points) when applied to test data from a different genre of that of the training set. Pradhan et al. (2008) showed that this perfor- mance degradation is essentially caused by the argu- ment classification subtask, and suggested the lexi- cal data sparseness as one of the main reasons. The same authors studied the contribution of the different feature types in SRL and concluded that the lexical features were the most salient features in argument classification (Pradhan et al., 2007). In recent work, we showed (Zapirain et al., 2009) how automatically generated selectional preferences (SP) for verbs were able to perform better than pure lexical features in a role classification experiment, disconnected from a full-fledged SRL system. SPs introduce semantic generalizations on the type of ar- guments preferred by the predicates and, thus, they are expected to improve results on infrequent and unknown words. The positive effect was especially relevant for out-of-domain data. In this paper we ad- vance (Zapirain et al., 2009) in two directions: (1) We learn separate SPs for prepositions and verbs, showing improvement over using SPs for verbs alone. (2) We integrate the information of several SP mod- els in a state-of-the-art SRL system (SwiRL1) and show significant improvements in SR classifica- tion. The key for the improvement lies in a meta- classifier, …

Content truncated to 500 words