Naslov
Word sense disambiguation : Distinguishing between individuals and kinds

Autori
Mikelić, Nives

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, magistarski rad

Fakultet
Computer Laboratory

Mjesto
Cambridge

Datum
30.07

Godina
2004

Stranica
60

Mentor
Copestake, Ann

Ključne riječi
Word sense disambiguation; British National Corpus; kind feature extractor; machine learning

Sažetak
This thesis presents a method of semi automatically acquiring kind properties of mass nouns, as well as nouns denoting broader domain of physical objects from the British National Corpus (BNC). Kind-referring nouns are not easily extracted, since there is no morphological distinction between kind reading of a noun and its individual reading and since there is no overall agreement in linguistic theories on the features that should be reliable clues for the distinction. In order to extract kind nouns together with their features from the BNC, the kind feature extractor was built. High reliable features were then used to semi automatically tag nouns as either KIND or PORTION in the first phase or as either KIND or OTHER in the later phase. The project results show that the method used is quite successful. Taking into account the rare number of kind-referring nouns in the corpus, features extracted with KFE seem to be highly reliable. Two different machine learning environments, TiMBL and WEKA, were used to learn these features and to predict the class of previously unseen words. Classifiers that were trained on the above mentioned features obtained accuracy of 89% and showed the ability to uncover some new patterns.

Izvorni jezik
Engleski

Znanstvena područja
Informacijske i komunikacijske znanosti

POVEZANOST RADA