The Duluth lexical sample systems in Senseval-3

The Duluth Lexical Sample Systems in SENSEVAL-3 Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN 55812 tpederse@d.umn.edu http://www.d.umn.edu/˜tpederse Abstract Two systems from the University of Minnesota, Duluth participated in various SENSEVAL-3 lexi- cal sample tasks. The supervised learning system is based on lexical features and bagged decision trees. It participated in lexical sample tasks for the English, Spanish, Catalan, Basque, Romanian and MultiLingual English-Hindi data. The unsuper- vised system uses measures of semantic relatedness to find the sense of the target word that is most re- lated to the senses of its neighbors. It participated in the English lexical sample task. 1 Introduction The Duluth systems participated in various lexical sample tasks in SENSEVAL-3, using both super- vised and unsupervised methodologies. The supervised lexical sample system that partic- ipated in SENSEVAL-3 is the Duluth3 (English) or Duluth8 (Spanish) system as used in SENSEVAL- 2 (Pedersen, 2001b). It has been renamed for SENSEVAL-3 as Duluth-xLSS, where x is a one let- ter abbreviation of the language to which it is be- ing applied, and LSS stands for Lexical Sample Su- pervised. The idea behind this system is to learn three bagged decision trees, one using unigram fea- tures, another using bigram features, and a third using co–occurrences with the target word as fea- tures. This system only uses surface lexical fea- tures, so it can be easily applied to a wide range of languages. For SENSEVAL-3 this system partici- pated in the English, Spanish, Basque, Catalan, Ro- manian, and MultiLingual (English-Hindi) tasks. The unsupervised lexical sample system is based on the SenseRelate algorithm (Patwardhan et al., 2003) for word sense disambiguation. It is known as Duluth-ELSU, for English Lexical Sample Un- supervised. This system relies on measures of semantic relatedness in order to determine which sense of a word is most related to the possible senses of nearby content words. This system de- termines relatedness based on information extracted from the lexical database WordNet using the Word- Net::Similarity package. In SENSEVAL-3 this sys- tem was restricted to English text, although in fu- ture it and the WordNet::Similarity package could be ported to WordNets in other languages. This paper continues by describing our super- vised learning technique which is based on the use of bagged decision trees, and then introduces the dictionary based unsupervised algorithm. We dis- cuss our results from SENSEVAL-3, and conclude with some ideas for future work. 2 Lexical Sample Supervised The Duluth-xLSS system creates an ensemble of three bagged decision trees, where each is based on a different set of features. A separate ensemble is learned for each word in the lexical sample, and only the training data that is associated with a par- ticular target word is used in creating the ensemble for that word. This approach is based on the premise that these different views of the training examples for a given target word will result in classifiers that make com- plementary errors, and that their combined perfor- mance will be better than …

Content truncated to 500 words