Acquisition of Large Scale Categorial Grammar Lexicons

A system is presented for inducing Categorial Grammar (CG) lexicons for natural language from either unannotated or minimally annotated corpora extracted from the Penn Treebank. A combination of symbolic and stochastic methods have been used to build a computationally e ective and psychologically plausible system, which learns linguistically useful lexicons. There are a variety of parameters in the system, including the corpus annotation used, the knowledge given to the learner and the weight given to the symbolic and stochastic methods. We present results from a set of experiments that investigate these parameters. The results also show that the system performs well even when compared with systems used for simpler problems.