The Basque Country University system: English and Basque tasks

The Basque Country University system: English and Basque tasks Eneko Agirre IXA NLP Group Basque Country University Donostia, Spain eneko@si.ehu.es David Martinez IXA NLP Group Basque Country University Donostia, Spain davidm@si.ehu.es Abstract Our group participated in the Basque and En- glish lexical sample tasks in Senseval-3. language-specific feature set was defined forA Basque. Four different learning algorithms were applied, and also a method that combined their outputs. Before submission, the performance of the methods was tested for each task on the Senseval-3 training data using cross validation. Finally, two systems were submitted for each language: the best single algorithm and the best ensemble. 1 Introduction Our group (BCU, Basque Country University), participated in the Basque and English lexical sample tasks in Senseval-3. We applied 4 differ- ent learning algorithms (Decision Lists, Naive Bayes, Vector Space Model, and Support Vector Machines), and also a method that combined their outputs. These algorithms were previously tested and tuned on the Senseval-2 data for En- glish. Before submission, the performance of the methods was tested for each task on the Senseval-3 training data using 10 fold cross val- idation. Finally, two systems were submitted for each language, the best single algorithm and the best ensemble in cross-validation. The main difference between the Basque and English systems was the feature set. A rich set of features was used for English, includ- ing syntactic dependencies and domain infor- mation, extracted with different tools, and also from external resources like WordNet Domains (Magnini and Cavagli´a, 2000). The features for Basque were different, as Basque is an agglu- tinative language, and syntactic information is given by inflectional suffixes. We tried to rep- resent this information in local features, relying on the analysis of a deep morphological analyzer developed in our group (Aduriz et al., 2000). In order to improve the performance of the al- gorithms, different smoothing techniques were tested on the English Senseval-2 lexical sam- ple data (Agirre and Martinez, 2004), and ap- plied to Senseval-3. These methods helped to obtain better estimations for the features, and to avoid the problem of 0 counts Decision Lists and Naive Bayes. This paper is organized as follows. The learn- ing algorithms are first introduced in Section 2, and Section 3 describes the features applied to each task. In Section 4, we present the exper- iments performed on training data before sub- mission; this section also covers the final config- uration of each algorithm, and the performance obtained on training data. Finally, the official results in Senseval-3 are presented and discussed in Section 5. 2 Learning Algorithms The algorithms presented in this section rely on features extracted from the context of the target word to make their decisions. The Decision List (DL) algorithm is de- scribed in (Yarowsky, 1995b). In this algorithm the sense with the highest weighted feature is se- lected, as shown below. We can avoid undeter- mined values by discarding features that have a 0 probability in the divisor. More sophisticated smoothing techniques have also been tried …

Content truncated to 500 words