R. Basili, M. T. Pazienza, and P. Velardi
Several research groups have been engaged with the problem of automatically acquiring selectional patterns for syntactic and semantic disambiguation from training corpora. More recently, few papers proposed probabilistic word association models to generalize cooccurrence patterns, in order to improve the coverage of the acquired knowledge. Though lately the devised probabilistic models became rather sophisticated, the evaluation of the acquired word clusters is rather disomogeneous and controversial. In this paper the experience made at the NLP laboratory of the University of Roma, Tor Vergata on the acquisition of selectional restrictions with different degrees of expressivity is described. The nature and the coverage of the acquired symbolic knowledge is studied in the light of several experiments of corpus-driven probability-based methods. The integration of the induced information with available on-line thesaurus (e.g. WordNet) is also analyzed.