X killed 
died
X lifted rose
X gave Z to received Z
X persuaded Y that 
came to
believe that P
X kill Y : X cause [Y die] X lift Y : X cause [Y rise] X give Z to Y : X cause [Y receive Z] X persuade Y that P : X cause [Y come to believe that P]
X cause [E to occur] 
E occurs
In ji20a, each verbal item is associated with a lexically specific inference rule. Each of these rules, however, is closely related to the others. Recasting the verbs in terms of a general function cause, as done in ji20b, allows one inference rule ji20c to be stated over the conceptual structure associated with each of these verbs.
Two questions which we must ask ourselves at this junction are, 1) Why do we want to represent such inferences in terms of lexical structure? and 2) Is it motivated to tailor our lexical structures in order to capture certain inferential distinctions between lexical items (i.e. should we postulate particular function-argument structures in order to capture semantic variations which may not have a syntactic effect)?
The answer to the first question is quite clear with reference to the example just discussed: specification of inferences which hold for elements of lexical structure which are common to many words allows an inference shared by these words to be captured in a general way. A lexical semantic relation (e.g. function-argument structure) with independent motivation from syntax which has a consistent inference associated with it thus serves an explanatory role in the specification of the semantic relationship within groups of words.
The second question is a bit more tricky. It is clear that there are often inferences which hold for groups of words semantically related in some way, while the way in which they are related cannot be shown to be a factor affecting syntax. Consider the distinction between the sentences in ji26, for example. The verbs drag and throw are related in that they both express the motion of something somewhere caused by someone's (Bill's) action. However, they differ in the temporal relations they express and in the entailments they have -- in ji26a, the causing action and the effected motion are temporarily coextensive and Bill ends up at the endpoint of the path, while in ji26b, the cause precedes and initiates the effect, and Bill does not move along the path.
Bill dragged the ball into the field (?and ran to the field to get it back). Bill threw the ball into the field (and ran to the field to get it back).
Would we wish to argue that the semantic element on which they differ (entraining vs. launching causation, in Jackendoff's terminology) should be captured in the lexical semantics of the words, since the inference affects what can subsequently be stated? If the answer is no, then we seem to be excluding the capture of a potential generalisation, for then there is no way to reflect the fact that words such as drag, tow, pull, etc. are related in some way and therefore no way to specify that they share an inference. Furthermore, the semantic difference between the verbs in ji26a and ji26b constrains what can felicitously follow the sentences containing these verbs in the discourse. The generalisation could only be captured if a rich non-linguistic knowledge representation were assumed which would express the semantic relationship between the words in question and allow the specification of inferences associated with such semantically related words. We would have to assume, for instance, a hierarchical knowledge representation in which inferences were inherited from supertypes. This hierarchy would have to be very carefully constructed in order to capture all the possible semantic relationships which can be shared among groups of words.
If the answer is yes, we are in danger of introducing non-linguistic
information into the lexicon and then the amount of information which
must potentially be included in the lexicon could not be
constrained and the types of inference procedures we would have
to have in the lexicon would have to be extremely powerful.
The problem is complicated by the fact that there is no
clear test -- not even ungrammaticality -- which can distinguish
between ``linguistic'' and ``non-linguistic'' information. Consider
the following example from copestake:92:
The rabbit bit John. *?The newborn rabbit bit John. *The rabbit with no teeth bit John.
Here we have a grammaticality difference which depends on arbitrary bits of world knowledge, e.g. that something must have teeth in order to bite, and that newborn rabbits have no teeth but rabbits in general do. We clearly would not wish to attempt to capture this semantic anomaly through lexical restrictions, or to claim that this knowledge is linguistic. There is no obvious test which can be used to distinguish between infelicities that depend on world knowledge and infelicities that can best be captured through lexical semantic structure and the constraints upon this structure (cf. the discussion of linguistic vs.\ world knowledge in Section 2.1).
The issues raised here really boil down to whether we should view semantic relationships between lexical items as lexical knowledge or world knowledge. The decision taken with respect to this question must be made very carefully and may have to be made on a case-by-case basis. Given that the main point of developing a lexical semantic framework is to capture generalisations about the way that language is used, it is worthwhile to reflect semantic relationships between words which have general application -- i.e. for which the inferences hold across many words and for which the semantic element involved in the relationship can be shown to influence the interpretation and use of several word classes. It seems particularly motivated to attempt to capture semantic elements lexically when the relationships they express hold across semantic domains.
In this section, I will review the ways in which Jackendoff's theory allows inferences to be captured. Although there may be other ways of capturing the inferences we would like to capture about sentences, it should become clear from the examples to be introduced below that Jackendoff's approach provides a clearly-defined framework within which inference patterns can be captured in regular ways. His work goes quite a long way towards identifying which relationships between words could be considered general and therefore which should be implemented lexically.