We have been developing a system for creating three dimensional computer graphics (3D-CG) that can accept natural-language sentences as an input, in addition to the mouse and pen-tablet inputs. Our system is unique in that when it does not know the shape of a noun used in the input sentence, it can estimate the shape by consulting the WordNet lexical database. In the present study, we have improved our system so that it can understand a wider variety of words and sentences than the previous system. We have also compared the performance of six different methods for the shape estimation.