Learning Greek Verb Complements: Addressing the Class Imbalance

Learning Greek Verb Complements: Addressing the Class Imbalance Katia Kermanidis, Manolis Maragoudakis, Nikos Fakotakis, George Kokkinakis Wire Communications Laboratory University of Patras, Rio, 26500, Greece {kerman,mmarag,fakotaki,gkokkin}@wcl.ee.upatras.gr Abstract Imbalanced training sets, where one class is heavily underrepresented compared to the oth- ers, have a bad effect on the classification of rare class instances. We apply One-sided Sam- pling for the first time to a lexical acquisition task (learning verb complements from Modern Greek corpora) to remove redundant and mis- leading training examples of verb non- dependents and thereby balance our training set. We experiment with well-known learning algorithms to classify new examples. Perform- ance improves up to 22% in recall and 15% in precision after balancing the dataset1. 1 Introduction Among the dependents of a verb, arguments are key participants in the event described by the verb, while adjuncts comprise secondary information concerning the ‘setting’ of the event (its context, location etc.). In previous work in automatic complement- adjunct distinction, Buchholz (1998) uses memory- based learning on the part-of-speech tagged and phrase structured part of the Wall Street Journal with a generalization accuracy of 91.6% and she includes verb subcategorization information in her data. Merlo and Leybold (2001) use decision trees to distinguish prepositional arguments from prepo- sitional modifiers. They incorporate semantic verb class, preposition and noun cluster information and reach an accuracy of 86.5% with a training set of 3692 and a test set of 400 instances. Aldezabal et al. (2002) work on Basque. They apply mutual in- formation and Fisher’s Exact Test to verb-case pairs (a case is any type of argument) which were obtained from a partially parsed newspaper corpus of 1.3 million words. Evaluation was performed by human tagging of the dependents of ten test verbs inside (55% f-measure) and outside (95% f- measure) the context of the sentence. Many re- searchers have attempted to distinguish comple- ments from adjuncts as a prerequisite for identifying verb subcategorization frames: Sarkar 1 This work was supported by the EU Project INSPIRE (IST- 2001-32746). and Zeman (2000) use a treebank and iteratively reduce the size of the candidate frame to filter out adjuncts. Briscoe and Carroll (1997) and Korhonen et al. (2000) use a grammar and a sophisticated parsing tool for argument-adjunct distinction. In this paper we address the issue of comple- ment-adjunct distinction in Modern Greek (MG) texts using well-known machine learning tech- niques (instance based learning, Naïve Bayes, and decision trees) and minimal resources. We make use of input that is automatically annotated only up to the phrase level, where the verb dependents are not identified. Therefore, a significant dispropor- tion between the number of complements and non- complements (adjuncts and non-dependents) arises among the candidates (complements being signifi- cantly fewer). This disproportion causes a signifi- cant drop in the minority (or positive) class (i.e. complements) prediction accuracy. Henceforth by adjuncts we will mean non-complements. The problem of class imbalance has been dealt with in previous work in different ways: oversampling of the minority class until it consists …

Content truncated to 500 words