Although I have argued that the heads first approach usually implies that the logical form of a node is instantiated at the time this node has to be generated it is possible to write reasonable grammars where this is not the case. For example raising-to-object constructions can be analysed in a way that is problematic for BUG1. Assume that the logical form for the sentence john believes mary to kiss the boy will be . Furthermore assume that is the syntactic object of . A reasonable definition of the lexical entry in the spirit of 2.1 then is the following:
node(vp( np/Np, [ np/Np2, vp(np/Np2,)/Vp ])/believes(Np,Vp), [believes|X]-X) ---> .
This lexical entry has two complements, a noun phrase and a verb phrase. Furthermore it is stated that the logical form of this noun phrase is identical to the logical form of the subject of the embedded verb phrase.
Such an analysis can be defended as follows. Assume that passive is accounted for by some lexical rule that, intuitively speaking, takes the object from the subcat list and makes it the subject (such as in LFG). Now if is defined as above then this passive rule will also naturally account for sentences such as mary is believed to kiss the boy.
Now, as the generator proceeds bottom-up it will try to generate the object noun phrase before the embedded verb phrase has been generated, i.e., before the link between the embedded subject and the object is found. As a result the logical form of the object is not yet instantiated and therefore BUG1 will not terminate in this case. Assuming that the analysis of raising-to-object is correct then it might be necessary to augment BUG1 with some version of goal freezing. If the generator comes across a uninstantiated logical form, then the execution of that node is suspended until the logical form is instantiated. In the case of this will imply that the embedded verb phrase will be generated first, after which the object can be generated. An implementation of this technique is reported in .