Influence of Conditional Independence Assumption on Verb Subcategorization Detection

Learning Bayesian Belief Networks from corpora has been applied to the automatic acquisition of verb subcategorization frames for Modern Greek (MG). We are incorporating minimal linguistic resources, i.e morphological tagging and phrase chunking, since a general-purpose syntactic parser for MG is currently unavailable. Comparative experimental results have been evaluated against Naive Bayes classification, which is based on the conditional independence assumption along with two widely used methods, Log-Likelihood (LLR) and Relative Frequencies Threshold (RFT). We have experimented with a balanced corpus in order to assure unbiased behaviour of the training model. Results have depicted that obtaining the inferential dependencies of the training data could lead to a precision improvement of about 4% compared to that of Naive Bayes and 7% compared to LLR and RFT Moreover, we have been able to achieve a precision exceeding 87% on the identification of subcategorization frames which are not known beforehand, while limited training data are proved to endow with satisfactory results.