UCD-S1: A hybrid model for detecting semantic relations between noun pairs in text

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 378–381, Prague, June 2007. c⃝2007 Association for Computational Linguistics UCD-S1: A hybrid model for detecting semantic relations between noun pairs in text Cristina Butnariu School of Computer Science and Informatics University College Dublin Belfield, Dublin 4, Ireland. ioana.butnariu@UCD.ie Tony Veale School of Computer Science and Informatics University College Dublin Belfield, Dublin 4, Ireland. tony.veale@UCD.ie Abstract We describe a supervised learning approach to categorizing inter-noun relations, based on Support Vector Machines, that builds a differ- ent classifier for each of seven semantic rela- tions. Each model uses the same learning strategy, while a simple voting procedure based on five trained discriminators with vari- ous blends of features determines the final categorization. The features that characterize each of the noun pairs are a blend of lexical- semantic categories extracted from WordNet and several flavors of syntactic patterns ex- tracted from various corpora, including Wikipedia and the WMTS corpus. 1 Introduction The SemEval task for classifying inter-noun semantic relations employs seven semantic relations that are not exhaustive: Cause-Effect, Instrument-Agency, Product-Producer Origin- Entity, Theme-Tool, Part-Whole and Content- Container. The task is to classify the relations between pairs of concepts that are part of the same syntactic structure in a given sentence. This approach employs a context-dependent classification, as opposed to usual out-of-context approaches in classifying semantic relations between noun pairs (e.g., (Turney, 2005), (Nastase et. al., 2006)). Our approach is based on the Support Vector Machines learning paradigm (Vapnik, 1995), in which supervised machine learning is used to find the most salient combination of features for each semantic relation. These features include semantic generalizations of the noun-senses as encoded as WordNet (WN) hyponyms, some manually selec- ted linguistic features (e.g., agentive, gerundive, etc.) as well as the observed relational behaviour of the given nouns in three different corpora: the col- lected glosses of WordNet; the collected text of Wikipedia; and the WMTS corpus. One can find similar approaches in the literature to the semantic classification of noun compounds. Turney (2005) uses automatically extracted para- phrases to build a similarity measure between pairs of concepts, while Nastase et. al. (2006) proposes separate models for two different word representa- tions when determining the semantic relation in modifier-noun compounds: a model based on the lexico-semantic aspects of words and a model that uses contextual information from corpora. Our ap- proach is different in that we use all the available features of word representations and concept inter- actions in a single hybrid model. 2 System description Our system, named the Semantic Relation Dis- criminator (or SRD), takes as input a set of noun pairs that are manually classified as positive/negat- ive for a given semantic relation and produces as output a discriminator for that semantic relation. We used SRD to learn different models for each of the seven semantic relations in the classification scheme for task 4 in the SemEval Workshop. The SRD system relies on several data-resources and tools: the WN noun-sense hierarchy, a corpus made …

Content truncated to 500 words