With an understanding of syntactic structure and thematic roles, we can venture into the world of trying to represent the meaning of a sentence. One method of representing meaning has evolved from the linguistics world. This representation is called "logical form". If, for example, we are given the sentence "John kissed Mary", the resulting parse tree is:
(S (SUBJ (NP (NAME John)))
(MAIN-V (VERB kissed))
(OBJ (NP (NAME Mary))))
and the corresponding logical form is:
(assert (past k1 kiss-action (agent k1 (name j1 person "John"))
(theme k1 (name m1 person "Mary"))))
What does it all mean? Essentially that the sentence is an assertion, in the past tense, in which the action is something called "kiss-action", the agent is a person whose name is "John", and the thematic object is a person named "Mary". The label k1 distinguishes this instance of the kiss-action from others that we might deal with, and j1 and m1 distinguish these persons from others with the same name.
The thing to note here is that logical form, although it's touted as a semantic representation, has a real strong syntactic flavor to it. It's very formal and precise, which makes it good for resolution techniques and so on. But the verb representation is weak---the logical form for "John kissed Mary" looks just like "John hit Mary" except that "kiss-action" becomes "hit-action". Intuitively, this just doesn't seem to capture the semantic difference between kissing someone and hitting someone. We want more out of the "kiss" than "kiss-action".
When we were talking about semantic representations last week, we came very close to talking about the decomposition of verbs into smaller component actions. We didn't actually get there, because I purposely steered you away to keep you on a different track. But it's an important idea, and this notion of decomposing events into combinations of smaller primitive acts is at the heart of another semantic representation scheme called "conceptual dependency".
The essence of conceptual dependency, or CD, representation is that understanding a sentence is more than just finding a representation of the sentence as it is read or heard---it is finding representations for the events that are expressed by the sentence, whether those events are stated explicitly or are only implied by the sentence.
CD events can be one of two kinds: physical or mental. Both kinds share a simple representational framework, regardless of the sentence that describes the event(s). Every event is represented by a "conceptualization" which consists of one of a limited set of primitive actions and a few roles:
(action (actor role-filler)
(object role-filler)
(to role-filler)
(from role-filler)
(instrument role-filler))
CD provides a set of primitive actions which, it is said, can be combined to represent any physical event that you could see take place on a movie screen. These primitive actions provide a means of obtaining similar representations for sentences with similar meanings but different forms. For example, consider the two sentences "John gave Mary a book" and "Mary took a book from John". Both sentences share the concept of transfer of possession, and we'd like our semantic representation to include that notion. Here's a first cut at how the interpretations of the two sentences might appear:
John gave Mary a book. Mary took a book from John.
(transfer-possession (transfer-possession
(actor John) (actor Mary)
(object book) (object book)
(to Mary) (to ?)
(from ?) (from John)
(instr ?)) (instr ?))
First, note that the only differences between "gave" and "took" are in the role-fillers for "actor", "to", and "from"--there's no difference in the action. Second, notice that we have no filler for the "from" role in the first sentence, no filler for the "to" role in the second sentence, and no filler for the "instr" role in either sentence. Can we fill these roles?
In "John gave Mary a book" the "to" role is explicitly mentioned in the sentence, but the "from" role is not. Yet we can be very sure that the book came from John. Could we possibly say "John gave Mary a book but it didn't come from John"? It's hard to see how that would be possible, so we can safely conclude that "John" fills the "from" role in the first sentence. Similarly, it's difficult to see how "Mary took a book from John but it didn't go to Mary" could be true. Hence, we can safely say that "Mary" fills the "to" role in the second sentence. While the sentences don't explicitly tell us what should fill those roles, our knowledge of the way the world and our language works tells us that we can fill those roles. We have to infer the fillers, but we know that they're right. Because we can't contradict those inferences using the "but test", we must include those inferences as part of our representation of the sentences.
But what about the "instr" roles? We can infer instruments for a transfer of possession, but no matter what we infer, we can think of plausible contradictions. For example, we might want to infer that John gave Mary a book by handing it to her. But is it possible to give the book to Mary without handing it to her? Sure, so that's not a mandatory inference. In some contexts it might be true, but not in all contexts, so we don't make it a required part of our representation for the sentence.
One other thing: in CD we don't call the transfer-of-possession act "transfer-of-possession". Historically, the CD folks have tried to some extent to distance the names of the primitive acts from from their related English verbs. So transfer of possession is denoted by ATRANS, which is short for abstract transfer. Here are the 11 primitive actions used in CD representations:
Abstract Acts:
ATRANS - abstract transfer; transfer of possession
Physical Acts:
PTRANS - physical transfer; change of location
PROPEL - apply force to an object (but not necessarily set in motion)
INGEST - take an object inside the body
Mental Acts:
MTRANS - mental transfer of information
MBUILD - construction of new information from old information
Instrumental Acts:
EXPEL - expulsion of object from the body
MOVEB - move body part
GRASP - grasp an object
ATTEND - focus a sense organ
SPEAK - make noise
Here are some examples of English sentences represented in CD form (but they're not the ones we looked at in class):
1) John gave Mary a book.
(atrans (actor John)
(object book)
(to Mary)
(from John)
(instr ?))
Since we don't know the mechanism by which the book was given, we can't fill the instrument role.
2) John gave Mary a punch.
(propel (actor John)
(object ? (part-of Mary))
(from ?)
(to ?)
(instr (propel (actor John)
(object fist (part-of John))
(to ? (part-of Mary))
(from John)
(instr (moveb (actor John)
(object fist (part-of John))
(to ? (part-of Mary))
(from John)))))
(result (phys-contact (object fist (part-of John))
(object ? (part-of Mary)))))
If we assume that John punched Mary, not that he gave her a drink, the above CD is the correct one. Note that we've added a slot called result. This indicates the state or states that are result from (or are caused by) the event. Although the sentence doesn't explicitly state that physical contact was made, it makes no sense to say "John gave Mary a punch but he never touched her." So, the "but test" tells us that the resulting state must be part of the semantic representation of "John gave Mary a punch." Here's an example of something that is inferred, but must be correct.
Acts result in changes of state, and states in turn enable new acts to be performed. There is no well-defined set of primitive states provided by the CD theory. There is a standard framework for representing states, though, in which the entity that has undergone some state change is named and the degree of change is indicated:
(state (object name) (value -10 to +10))
We more or less invent "state" as we need it. Some examples of states with a range of values are:
health -10 (dead) to +10 (perfect health)
fear -10 (terrified) to 0 (calm)
anger -10 (furious) to 0 (calm)
joy -10 (hopelessly sad) to +10 (ecstatic)
Other states, like phys-contact above, are binary in nature, and don't use a variable scale.
3) John gave Mary a call.
If we assume that John called out to Mary, the CD would look like this:
(speak (actor John)
(object ?)
(to Mary)
(from John))
If we assume that Mary heard John and some information was passed, the CD looks like this:
(mtrans (actor John)
(mobj ?)
(to cp (part-of Mary))
(from cp (part-of John))
(instr (speak (actor John)
(object ?)
(to Mary)
(from John))))
Note that what used to be the "object" role is now labeled "mobj". That stands for "mental object"...it's a notation I always forget to use when I talk about this stuff, but using it helps you remember that mental objects are different than physical objects. In CD theory, all mental objects must be conceptualizations. So if we knew what John said, we'd represent it as one or more events and replace the question mark after "mobj" with that structure.
And then if we want to assume that a telephone was involved, the CD will be very complicated. What follows is a reasonable first approximation, and completely acceptable for our purposes, but hardcore CD types would probably go into much more detail than this:
(mtrans (actor John)
(mobj ?)
(to cp (part-of Mary))
(from cp (part-of John))
(instr (speak (actor John)
(object ?)
(to Mary)
(from John))
(instr (speak (actor John)
(object ?)
(to telephone)
(from John)))))
This assumes that somebody picked up the phone, because John wouldn't say anything if nobody answered. And it assumes that the person on the other end listened, which is not represented here. We could make this really messy, but I'll spare you all that.
4) John gave Mary a headache.
(do (actor John)
(object ?)
(to ?)
(from ?)
(instr ?)
(result (phys-health (object Mary) (value -1))))
We don't know what John did, who he did it to, or anything else except that somehow Mary's physical health declined slightly because of it. Sure, (value -1) doesn't necessarily distinguish between a headache and some other minor ailment, but it will be good enough for now. If you encounter a situation where you want something better, you can add enhancements as needed. These CD reps aren't like the immutable laws of physics; they're just one guy's ideas about how to represent meaning. You can bend them within reason. Just be consistent and, if it's something you're being graded on like in this course, explain your reasoning.
Notice that all four sentences above have "gave" as the main verb. Also notice that the resulting CDs are very different from each other. That's what different meanings of the same verb should do for you. On the other hand, the logical form representations for these sentences would look pretty much the same. Ugh.
What would a logical form representation look like for sentence 4 above? The action would probably be something like "give-action27", "headache" would be treated like some tangible physical object, and we wouldn't have anything that told us that Mary's health declined ever so slightly. If you compare that to what the corresponding CD tells us, it becomes fairly clear that CD is going to be a much more useful tool for getting computers to understand natural language and not just process it.
Also, CD gives us one more thing that LF doesn't. It explicitly points out what is not known in the semantic representation for some sentence. For example, in the "headache" example, the LF indicates that we know what John did, but the CD points out that we don't have a clue. We only know what the result was. That's an important difference, and we can take advantage of it when we try to connect individual sentences together in a coherent story.
That's not to say that CD is perfect. Can we represent everything we really need to with combinations of just 11 primitive actions? It can be difficult. Try "The man who looked like Cary Grant smiled knowingly." Ouch. How do we know we have the right primitives? Are there different sets of primitives for different cultures or contexts? Do dolphins have an ATRANS?
There are other, competing sets of primitives that are offered up as "conceptual" and/or "interlingual" representations, but so far they all suffer from the same problems to varying degrees. Nobody has found the perfect solution yet, so you can still write that doctoral dissertation if you're so inclined.
Finally, notice that if we change the value from -1 to -10 in sentence 4 above, we now have the CD representation for "John killed Mary." On that happy note, we'll move to conceptual analysis.
Copyright (c) 2004 by Kurt Eiselt. All rights reserved, except as previously noted.
Last revised: March 30, 2004