Wordnet-Based Cross-Language Identification of Semantic Relations
Ivelina Stoyanova, Svetla Koeva, Svetlozara I. Leseva · 2013
We propose a method for cross-language identification of semantic relations based on word similarity measurement and morphosemantic relations in WordNet. We transfer these relations to pairs of derivationally unrelated words and train a model for automatic classification of new instances of (morpho)semantic relations in context based on the existing ones and the general semantic classes of collocated verb and noun senses. Our experiments are based on Bulgarian-English parallel and comparable texts but the method is to a great extent language-independent and particularly suited to less-resourced languages, since it does not need parsed or semantically annotated data. The application of the method leads to an increase in the number of discovered semantic relations by 58.35 % and performs relatively consistently, with a small decrease in precision between the baseline (based on morphosemantic relations identified in wordnet) – 0.774, and the extended method (based on the data obtained through machine learning) – 0.721. 1