Using conceptual dependency: Write CD representations for the following sentences. Unfilled slots in the CD representations can either be left in (to make the missing information more obvious) or omitted for convenience (CD-based programs would always know they're unfilled, though). Because we will not address object representations, try to make up your own scheme for representing objects, analogous to CD's representation for actions. If you notice an ambiguity that must be resolved before the representation can be done, point out the ambiguity and how you are resolving it.
Sally drove a motorcycle to New York.
(ptrans (actor Sally) (object motorcycle) (to New York))
John donated his sofa to the Salvation Army.
(atrans (actor John) (object sofa) (from John) (to Salvation Army))
John roared into town in a new BMW.
(ptrans (actor John) (object BMW) (to town))
John ate popcorn with his foot.
(ingest (object popcorn) (actor John) (insturment (ptrans (actor John) (object foot) (to mouth))))
Mary broke John's TV.
(broke (actor Mary) (object TV))
Ambiguity of words: As the previous example shows, MicroELI can handle ambiguous verbs fairly easily. Discuss the issues involved in writing requests that would enable mcELI to handle ambiguous nouns. (E.g., the word ``bill'' in ``Bill paid the bill.'') Sketch how you might change the MicroELI framework and how you would write requests using the new framework to enable the program to parse the sentence ``Bill paid the bill with a bill.''
The main problem is the fact that the meaning of the words must be distungushed by referencing their location in the sentence. However, since sentence contructions are so varied, writing a parser to consider all posibilities would be difficult.
In order to parse the sentence correctly, ELI would look at the relation of the word to the rest of the sentence. For example, if the word is found to be the subject of the sentence and is singular it must be a proper-noun refering to a person. Should it be the object of a prepesition, one could assume that it is a bill describing a monetary amount. If it were preceded with an article, it would be a bill of amounts due.
To add this functionality to ELI the ability to destingush between singular and plural nouns must be added. Extendeding the sentence contruct recognition would also be neccisary in order to find prepisitional phrases.
Request to the system would not change as ELI could parse the sentence as:
Bill (noun-subject-person) paid (verb) the bill (noun-dues) with a bill (prepisitional phrase-money).
Difficulty of CD: Spend some time thinking about what makes a sentence hard or easy to represent using CD. What kinds of sentences are easy to represent? What kinds of sentences are hard? Give concrete examples of sentences illustrating the types of issues you identify. Why?
Hard sentences would be any ones in passive voice (The car he drove crashed) or, as mentioned above, ones with the same word taking on similar meanings (Till the ground till Monday), and ones containing words with multiple meanings that can only be understood from previous context (He set the clock on the desk). These are difficult becuase they all take advantage of a person's ability to generalize the solution based on context.
Easy phrases would be sentences using the main 12 abilities (He went home) and ones with a small number of interpritations (He bought some crack). These are easy since they do not require many complicated rules to define how to understand them. If a word has less meanings it has less of a chance of being parsed incorrectly. Also if the verb is well defined (not like "grabbed a bus", "grabbed a bite", "grabbed the ecstacy") then it can be easily out into 1 of the 12 groups.
A Problem: Assume that the sentence ``A coin rolled into the store'' is analyzed as (PTRANS (OBJECT (MONEY)) (TO (STORE))) Process this line with micro-SAM. Then process the CD representation for ``John bought a restaurant.''
Explain what goes wrong in each case and the causes of the problems.
What is a simple way that you could modify the code to address this? (You do not need to write this code; simply give your strategy.)
When might your simple approach fail?
What would you propose to address this problem? Try to think of a method that is robust (applies to a wide range of situations) and efficient, and sketch its tradeoffs. (Addressing this problem in a rich way has been the subject of much work in the literature, as well as the heart of at least one dissertation in AI.)
((ptrans (actor (money)) (object (money)) (to (store)))
(shopping (shopper (money)) (store (store))))
The parser sees the object 'money' and the to 'store' and assumes it is dealing with a shopping story.
((ptrans
(actor (person (name (jack))))
(object (person (name (jack))))
(to #f))
(mtrans
(actor (person (name (jack))))
(object (restaurant))
(to waiter)
(from (person (name (jack)))))
(atrans
(actor waiter)
(object (restaurant))
(from waiter)
(to (person (name (jack)))))
(eating
(order (restaurant))
(eater (person (name (jack))))))
The parsers sees the restaurant and a person atransing something. Based on that info it things this is a eating story.
One way to fix this is to make the links to scripts more specific. Including more data in the script so that a simple sentence will not get identified as that story. This will fix general sentences getting interpreted as the wrong type, but will also make it harder for valid sentences to be identified.
A more robust solution would be to define more scripts so that general sentences will be identified correctly and not misinterpreted as other stories. This would require adding a large number scripts to catch most general sentences.