It has often been argued that lexicalized grammars are better to process bottom-up than top-down because the lexicon defines important information for termination and because lexicalized grammars like Categorial Grammar or only allow very general predictions because of their use of very general rule schemata. Clearly, the first aspect is valid, if only simple top-down methods are used. However, using complex (and clever) strategies, these termination problems are avoided.
If only general rule schemata are used, we can say the following. Because of the general nature of the schematic rules, and because our algorithm tries to look for lexical information as early as possible, the size of a predicted derivation tree is small (in the case of lexical grammars). Only when lexical information is available, its size will increase but be limited with respect to the subcategorization information of the lexical entry. However, if lexical entries are identified for the next parts of the input then more and more constraints for constructing the possible analysis are spread over the derivation tree. But then, more specific predictions are possible, and the algorithm is strongly goal-directed even in the case of lexicalized grammars like .