Üôðððòòòò Ûý Ñùùøýý Äääöòòòò Úöö ××ðð Blockinøøóòòð Ôööööööò

with Bayesian networks Massimiliano Ciaramita and Mark Johnson Cognitive and Linguisti S ien es Box 1978, Brown University Providen e, RI 02912, USA massimiliano iaramita brown.edu mj s.brown.edu Abstra t This paper presents a Bayesian model for unsupervised learning of verb sele tional preferen es. For ea h verb the model reates a Bayesian network whose ar hite ture is determined by the lexi al hierar hy of Wordnet and whose parameters are estimated from a list of verbobje t pairs found from a orpus. \Explaining away", a well-known property of Bayesian networks, helps the model deal in a natural fashion with word sense ambiguity in the training data. On a word sense disambiguation test our model performed better than other state of the art systems for unsupervised learning of sele tional preferen es. Computational omplexity problems, ways of improving this approa h and methods for implementing \explaining away" in other graphi al frameworks are dis ussed. 1 Sele tional preferen e and sense ambiguity Regularities of a verb with respe t to the semanti lass of its arguments (subje t, obje t and indire t obje t) are alled sele tional preferen es (SP) (Katz and Fodor, 1964; Chomsky, 1965; Johnson-Laird, 1983). The verb pilot arries the information that its obje t will likely be some kind of vehi le; subje ts of the verb think tend to be human; and subje ts of the verb bark tend to be dogs. For the sake of simpli ity we will fo us on the verb-obje t relation although the te hniques we will des ribe an be applied to other verb-argument pairs. We would like to thank the Brown Laboratory for Linguisti Information Pro essing; Thomas Hofmann; Elie Bienensto k; Philip Resnik, who provided us with training and test data; and Daniel Gar ia for his help with the SMILE library of lasses for Bayesian networks that we used for our experiments. This resear h was supported by NSF awards 9720368, 9870676 and 9812169. ENTITY something FOOD LIQUID PHSYSICAL_OBJECT BEVERAGE LAND earth ISLAND drink COFFEE JAVA-1 JAVA-2 java ESPRESSO BALI island espresso bali liquid WATER ALIMENT TEA beverage entity ORGANISM