Evaluating lexicographer controlled semi-automatic word sense disambiguation method in a large scale experiment

Word Sense Disambiguation in text remains a difficult problem as the best supervised methods require laborious and costly manual preparation of training data. On the other hand, the unsupervised methods yield significantly lower precision and produce results that are not satisfying for many applications. Recently, an algorithm based on weakly-supervised learning for WSD called Lexicographer-Controlled Semi-automatic Sense Disambiguation (LexCSD) was proposed. The method is based on clustering of text snippets including words in focus. For each cluster we find a core, which is labelled with a word sense by a human, and is used to produce a classifier. Classifiers, constructed for each word separately, are applied to text. The goal of this work is to evaluate LexCSD trained on large volume of untagged text. A comparison showed that the approach is better than most frequent sense baseline in most cases.