Knowing a word by the grammatical company it keeps
European Media Laboratory GmbH, Heidelberg
Le 2 mai 2007 à 11h30
Salle 3195, Pavillon André-Aisenstadt
Word context, usually in the form of simple or skip n-grams, is actively used for building language models, classifying texts, and other NLP tasks. From a semantic point of view, n-grams are noisy, capturing word or expression sequences that are not actually related. An alternative is to use grammatically-motivated collocations, which reflect on a surface level deeper semantic relations. The description of a word through its grammatically motivated collocations can be seen as a rough approximation of its semantics. In this talk I will present experiments that show how well such a description performs for semantic analysis tasks such as recognizing semantic relations between noun-modifier pairs and word sense disambiguation.