FML-based SCF predefinition learning for Chinese verbs

This paper describes the first attempt to acquire Chinese SCFs automatically and the application of Flexible Maximum Likelihood (FML), a variational filtering method of the simple maximum likelihood (ML) estimate from observed relative frequencies, to the task of predefining a basic SCF set for Chinese verb subcategorization acquisition. By setting a flexible threshold for SCF probability distributions over 1774 Chinese verbs, we obtained 141 basic SCFs with a reasonably practical coverage of 98.64% over 43,000 Chinese sentences. After complementation of 11 manually observed SCFs, a both linguistically and intuitively acceptable basic SCF set was predefined for future SCF acquisition work.