The Johns Hopkins SENSEVAL2 system descriptions
David Yarowsky, Silviu Cucerzan, Radu Florian, Charles Schafer, Richard Wicentowski · 2001
This article describes the Johns Hopkins University ( JHU) sense disambiguation systems that participated in seven SENSEVAL2 tasks: four supervised lexical choice systems (Basque, English, Spanish, Swedish), one unsupervised lexical choice system (Italian) and two supervised all-words systems (Czech, Estonian). The common core supervised system utilizes voting-based classifier combination over several diverse systems, including decision lists (Yarowsky, 2000), a cosine-based vector model and two Bayesian classifiers. The classifiers employed a rich set of features, including words, lemmas and part-of-speech informatino modeled in several syntactic relationships (e.g. verb-object), bag-of-words context and local collocational n-grams. The allwords systems relied heavily on morphological analysis in the two highly inflected languages. The unsupervised Italian system was a hierarchical class model using the Italian WordNet. 1 The Feature Space The JHU SENSEVAL2 systems utilized a rich feature space based on raw words, lemmas and partof-speech (POS) tags in a variety of positional relationships to the target word. These positions include traditional bag-of-word context, local bigram and trigram collocations and several syntactic relationships based on predicate-argument structure (described in Section 1.2). Their use is illustrated on a sample English sentence for train in Figure 1. 1.1 Part-of-Speech Tagging and Lemmatization Part-of-speech tagger availability varied across the languages included in this sense-disambiguation system evaluation. Transformation-based taggers (Ngai and Florian, 2001) were trained on standard data