 |
Extracting Dutch Hypernym Pairs from the Web
Erik Tjong Kim Sang
University of Amsterdam
We apply pattern-based methods for collecting evidence for hypernym
relations between nouns. We vary the type of evidence, either for
cousin relations or hypernym relations and examine two types of data
sources: a large text corpus and the web. Additionally, we mine thousands
of hypernym prediction patterns from the corpus and investigate methods
for combining their predicted hypernym-hyponym pairs. We evaluate the
various approaches comparing their hypernym suggestion with the Dutch part
of EuroWordNet. We show that the abundance of available data enables the
web-based techniques to outperform the corpus-based techniques and obtain
reasonable results in predicting candidate hypernyms (precision: 41% and
recall 22%).
|
