Semi-supervised word sense disambiguation based on weakly controlled sense induction

Word Sense Disambiguation in text is still a difficult problem as the best supervised methods require laborious and costly manual preparation of training data. On the other hand, the unsupervised methods express significantly lower accuracy and produce results that are not satisfying for many application. The goal of this work is to develop a model of Word Sense Disambiguation which minimises the amount of the required human intervention, but still assigns senses that come from a manually created lexical semantics resource, i.e., a wordnet. The proposed method is based on clustering text snippets including words in focus. Next, for each cluster we found a core, the core is labelled with a word sense by a human and finally is used to produce a classifier. Classifiers, constructed for each word separately, are applied to text. A performed comparison showed that the approach is close in its precision to a fully supervised one tested on the same data for Polish, and is much better than a baseline of the most frequent sense selection. Possible ways for overcoming the limited coverage of the approach are also discussed in the paper.