ILK: Machine learning of semantic relations with shallow features and almost no data

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 187–190, Prague, June 2007. c⃝2007 Association for Computational Linguistics ILK: Machine learning of semantic relations with shallow features and almost no data Iris Hendrickx CNTS / Language Technology Group Uversity of Antwerp, Universiteitsplein 1 2610 Wilrijk, Belgium iris.hendrickx@ua.ac.be Roser Morante, Caroline Sporleder, Antal van den Bosch ILK / Communication and Information Sciences Tilburg University, P.O. Box 90153, 5000 LE Tilburg, The Netherlands {R.Morante,C.Sporleder, Antal.vdnBosch}@uvt.nl Abstract This paper summarizes our approach to the Semeval 2007 shared task on “Classifica- tion of Semantic Relations between Nom- inals”. Our overall strategy is to develop machine-learning classifiers making use of a few easily computable and effective fea- tures, selected independently for each clas- sifier in wrapper experiments. We train two types of classifiers for each of the seven re- lations: with and without WordNet informa- tion. 1 Introduction We interpret the task of determining semantic rela- tions between nominals as a classification problem that can be solved, per relation, by machine learning algorithms. We aim at using straightforward features that are easy to compute and relevant to preferably all of the seven relations central to the task. The starting conditions of the task provide us with a very small amount of training data, which further stresses the need for robust, generalizable features, that generalize beyond surface words. We there- fore hypothesize that generic information on the lex- ical semantics of the entities involved in the rela- tion is crucial. We developed two systems, based on two sources of semantic information. Since the entities in the provided data were word-sense dis- ambiguated, an obvious way to model their lexical semantics was by utilizing WordNet3.0 (Fellbaum, 1998) (WN). One of the systems followed this route. We also entered a second system, which did not rely on WN but instead made use of automatically generated semantic clusters (Decadt and Daelemans, 2004) to model the semantic classes of the entities. For both systems we trained seven binary clas- sifiers; one for each relation. From a pool of eas- ily computable features, we selected feature subsets for each classifier in a number of wrapper exper- iments, i.e. repeated cross-validation experiments on the training set to test out subset selections sys- tematically. Along with feature subsets we also chose the machine-learning method independently for each classifier. Section 2 presents the system description, Sec- tion 3, the results, and Section 4, the conclusions. 2 System Description The development of the system consists of a prepro- cessing phase to extract the features, and the classi- fication phase. 2.1 Preprocessing Each sentence is preprocessed automatically in the following steps. First, the sentence is tokenized with a rule-based tokenizer. Next a part-of-speech tag- ger and text chunker that use the memory-based tag- ger MBT (Daelemans et al., 1996) produces part- of-speech tags and NP chunk labels for each token. Then a memory-based shallow parser predicts gram- matical relations between verbs and NP chunks such as subject, object or modifier (Buchholz, 2002). The tagger, …

Content truncated to 500 words