Machine Learning with Lexical Features: The Duluth Approach to SENSEVAL-2

Machine Learning with Lexical Features: The Duluth Approach to Senseval-2 Ted Pedersen Department of Computer Science University of Minnesota, Duluth Duluth, MN 55812 USA tpederse©d.umn.edu Abstract This paper describes the sixteen Duluth entries in the SENSEVAL-2 comparative exercise among word sense disambiguation systems. There were eight pairs of Duluth systems entered in the Spanish and English lexical sample tasks. These are all based on standard machine learning algo- rithms that induce classifiers from sense-tagged training text where the context in which am- biguous words occur are represented by simple lexical features. These are highly portable, ro- bust methods that can serve as a foundation for more tailored approaches. 1 Introduction The Duluth systems in SENSEVAL-2 take a su- pervised learning approach to the Spanish and English lexical sample tasks. They learn deci- sion trees and Naive Bayesian classifiers from sense-tagged training examples where the con- text in which an ambiguous word occurs is rep- resented by lexical features. These include uni- grams and bigrams that occur anywhere in the context, and co-occurrences within just a few words of the target word. These are the only types of features used. There are no syntac- tic features, nor is the structure or content of WordNet employed. As a result these systems are highly portable, and can serve as a founda- tion for systems that are tailored to particular languages and sense inventories. The word sense disambiguation literature provides ample evidence that many different kinds of features contribute to the resolution of word meaning. These include part-of-speech, morphology, verb-object relationships, selec- tional restrictions, lexical features, etc. When used in combination it is often unclear to what degree each type of feature contributes to over- all performance. It is also unclear to what 139 extent adding new features allows for the dis- ambiguation of previously unresolvable test in- stances. One of the long term objectives of our research is to determine which types of features are complementary and cover increasing num- bers of test instances as they are added to a representation of context. 2 Experimental Methodology The training and test data for the English and Spanish lexical sample tasks is split into sep- arate training and test files per word. A su- pervised learning algorithm induces a classifier from the training examples for a word, which is then used to assign sense tags to the test in- stances for that word. The context in which an ambiguous word oc- curs is represented by lexical features that are identified using the Bigram Statistics Package (BSP) version 0.4. This is free software that extracts unigrams and bigrams from text us- ing a variety of statistical methods. Each uni- gram or bigram that is identified in the training data is treated as a binary feature that indicates whether or not it occurs in the context of the word being disambiguated. The free software package SenseTools (version 0.1) converts train- ing and test data into a feature vector repre- sentation, based on the output from BSP. This …

Content truncated to 500 words