Understanding Complex Situations
Proceedings of the Sixth International Joint Conference on Artificial Intelligence. .
© 1979 International Joint Conferences on Artificial Intelligence
Abstract
Natural language texts often refer to complex situations. Many of these situations involve relationsh ips between people's goals. In order to build a program that understands texts, it is necessary to give the program knowledge about goal relationships and the situations to which they give rise. This knowledge constitutes a theory of planning about real world situations. We have incorporated this theory of planning into a program called PAM (Plan Applier Mechanism). As a result we now have a natural language processing system that can comprehend many complicated dramatic situations.
- Bruce, B. and Newman, D. (1978) Interacting Plans. In Co nitive Science, vol. 2, no. 3.
- Charniak, E. 1978 . On the use of framed knowledge in language comprehension. To appear in Artificial Intelligence.
- Cullingford, R.E . (1978). Script applica- tion: Computer understanding of newspaper stories. Yale Univ. Research Report #116.
- Sacerdoti, E. (1977). A Structure for Plans and Behavior, Elsevier, Amsterdam.
- Schank, R.C . and Abelson, R.P. (1977), Scripts, Plans, Goals and Understanding, Lawrence Erlbaum Press, Hillsdale, N.J.
- Wilensky, R. (l978a). Understanding goal- based stories . Yale Technical Report 140.
- Wilensky, R. (l978b). Why John married Mary: Understanding stories involving recurring goals. In Cognitive Science, vol. 2, no. 3. Yale Technical Report.
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
1. INTRODUCTION
A number of recent attempts have been made to address the problem of inference proliferation in natural language text understanding (for example, see Cullingford (1978) and Charniak (1978)). The thrust of this research has been to use knowledge to guide the inference pro cess. By equipping an understander with know ledge of the particular situation referred to in the text, the understander makes those in ferences which are the most plausible without generating many spurious or irrelevant ones.
The drawback of this approach is that a natural language understander cannot be expected to have familiarity with every situation it is going to encounter. A human reader can under stand many natural language texts that contain some novel elements in them. To build a natu ral language understanding system that is capable of understanding such text, a more general inferential capability is needed.
1.1 PAM
PAM (Plan Applier Mechanism) is a program that uses knowledge about people's intentions to understand situations that do not necessarily conform to stereotypical sequences of events. PAM applies knowledge about plans and goals to find explanations for the events in a text. That is, PAM explains a character's actions by inferring the plan that the action is a part of, and inferring the goal that the plan is aimed at achieving. Once a character's motivation has been determined, PAM uses this inference to help interpret subsequent events.
*This work was supported in part by the Advanced Research Projects Agency of the De partment of Defense and monitored under the Office of Naval Research under contract N00014-75-C-llll.
1.2 A Problem
The intentional knowledge originally given to PAM is based on the plan representation scheme of Schank and Abelson (1977). This scheme in cludes a number of important goals that a char acter could use to achieve a goal. PAM organ izes its knowledge about these entities so that it could infer a character's intentions while reading a story.
While this idea worked well for understanding simple stories, most interesting story situa tions do not conform to this notion of planning. Real stories tend to be about dramatic situa tions in which some difficulty is presented for a character. Usually, this difficulty involves a situation in which numerous goals are present at once. These goals can interact in compli cated ways, and knowledge about the nature of the ~relationships is needed to understand each character's behavior.
~e have been able to formalize some of the knowledge about goal relationships needed for story understanding. By supplying PAr1 with this knowledge, the program is now able to understand and answer questions about many com plicated stories that are beyond its former capac ity, as well as the capacity of other natural language processing systems.
The goal relationships that appear to play a central role in story situations are the following:
- Goal subsumption - A way of planning for recurring goals.
- Goal conflict- An inimicable relationship between simultaneous goals of the same character.
- Goal competition - An inimicable relation ship between simultaneous goals of dif ferent characters.
- {4) Goal concord - A mutually advantageous relationship between simultaneous goals of different characters.
Goal subsumption is discussed in detail in l~ilensky (1978b), and will only be touched on here. Instead, we shall be concerned with goal conflict and competition. A more detailed analysis of these relationships, as well as an ana lysis of ~oal concord, ·is found in Wil ensky (1978a).
2. GOAL SUBSUI1PTION
Goal subsumption is a form of planning in which a character plans in anticipation of having a set of goals, rather than in reaction to each indivi dual goal. For example, consider the foll owing story:
- John got a job in another town. He asked Fred if he would sell him his car.
Instead of having a single goal, here John has a goa l that is likely to recur repeatedly. Since having a job gives rise to the goal of being at work each day, it is possible to plan for these recurring goals at once rather than dealing with each one as it arises. John's pl an to purchase a car makes it easier to achieve each of these goals since he would no longer have to obtain possession of a vehicle on each occasion.
Goal subsumption is composed of the following components:
- A set of anticipated goals.
- Plans for each goal, each plan sharing a common precondition.
- A state that meets this precondition and
endures over a period of time during which the qoals are expected to arise (called a goal subsumption state).
2.1 Goal Subsumption Situations
Goal subsumption is a useful concept in story understanding because there are a number of frequently occurring story situations based on it. The most important of these situations occurs when a functioning goal subsumption state is abruptly terminated, and is called Goal Subsumption Termination. The following is an example of a story PAr! can process that in volves such a situation.
PAll Ex amp 1 e:
ItiPUT: JOHrl AND t1ARY !·/ERE HAPPILY t1ARRIED. THEtl ONE DAY, JOHII \~AS KILLED IN A CAR ACCIDENT. t1ARY HAD TO FIND .A. JOB .
Question:
ItiPUT: HHY DID 11ARY NEED El1PLOYr1EtH?
RESPONSE : JOHN DIED AND SO SHE NEEDED A SOURCE OF INCOt1E.
PN~ infers that John's death terminates a sub sumption state for l1ary, and that she may seek to replace it. Since traditional marriage sub sumes certain ecomonic goals that can also be subsumed by having a job, PM1 infers that this is_the_reason behind r1ary's goal. This explan atlon 1s accessed and given as the answer to the question above.
3. GOAL CONFLICT
Consider the following situations:
- John wanted to go bowling with his friends , but he had rromi sed t1ary he wou 1 d take her out to dinner that evening.
- John wanted to live in San Francisco, but he also wanted to live near his relatives in Kansas.
- John wanted to go to the football game, but it was raining outside.
In each of these stories, a character has a number of goals that interfere with one anoth~r. Such a situation is called a~ confl1ct. To understand a story involving a ~oal conflict, a story understander must be capable of detecting the presence of the con flict, and of using knowledge about the types of situations to which goal conflict may give rise. For example, suppose story (2) were continued as follows:
John decided to go bowling anyway. ~Jhen he got home, rlary to 1 d John she was going to divorce him.
To understand that 11ary wanted to divorce John because he had neglected her, a reader must in fer that John didn't take her out to dinner. This inference is based on the fact that John's goal of takjng ~1ary out was in conflict with another goal, and that this goal was pursued. \~hen two goa 1 s are in conflict, and one of them is pursued, a reader can infer that the other has been abandoned. This knowledge about goal conflict is required to find an explanation for ilary' s behavior.
3.1 ~of�oal Conflict
A classification of goal conflicts is needed to organize knowledge for both goal conflict de tection as well as for understanding the situa tions to which goal conflicts give rise. We have found the following taxonomy of goal con flicts to be useful:
Goal Conflict Classification
- ( 1 ) Goal conflicts based on resource limitations.
- Goa 1 conflicts based on mutually exclusive states.
- ( 3) Goal conflicts based on generating preservation goals.
Each of these classes is now described.
3. 1 . 1 Resource Limitations
Goal conflicts based on limited resources can come about in a number of 1~ays, depending upon the nature of the resource involved. Two im portant categories of resource limitations are found to exist:
- (a) Shortages based on time limitatjons
- (b) Shortages based on consumable resource limitations.
Due to a conflict based on a shortage of the consumable resource paper, we cannot afford a description of that kind of conflict here. He will discuss time limitations instead.
3.1.1.1 Time
Consider the following story:
- John wanted to watch the football game, but he had a paper due the next day. John watched the football gane. John failed Civics.
Story (5) is an instance of a goal conflict based on a shortage of time. In order for goals to conflict due to time resources, the plans for the goals must have some time res trictions on them; the amount of time avail able to the olanner must be less than that needed for the successful execution of the plans, and the planner must be incapable of carrying out the plans si~ultaneously. For ex ample, John must synchronize his plan of watch ing the football game with the game's broad cast; he must initiate his plan of writing a paper sufficiently before its deadline, and he is probably incapable of executing both plans at once because they each demand his attention .
Thus a goal conflict based on a limitation of time resources really also involves a limita tion of a functional object or of a person's abilities. However, the text of a story does not usually make all these limitations explicit. Instead, a reader must infer the conflict from hints in the text. Time-based goal conflicts are hinted at by one of the following occur ences in the text of the story:
- A statement of a deadline.
- A statement of a li~itation of a function al object.
- A statement of a limitation of a person's abi 1 ities.
- An explicit statement of a time shortage.
Hhen one of these statements occurs in a story, the reader can infer that a time shortage exists. For example, story (5) above contains an explicit statement of a deadline. Thus the reader infers the goal conflict from the state ment that John's paper' was due at a certain time.
This inference can be made in PAM with the following rule:
Ru 1 e T 1 :
If
a character has two or more goals, and
the story specifies that the plan for one of the goals must be initiated, completed, or occurring at or before a particular point in time,
then
infer that a goal conflict based on a shortage of time exists among that character's goals.
Hints (2), (3) and (4) are applicable to the following stories, respectively:
- John wanted to simmer two dishes on his camp stove, but the stove had only one burner.
- John had to write a paper for a course, and read a book for an exam, but John was a very slow reader.
- John wanted to visit both Fred and Bill, but he didn't have enough time to see them both.
The rules needed to infer the conflict hinted at i n each of these examples can be formulated similarly to rule T1 above.
Nhen a reader sees one of these hints, the reader infers that a goal conflict is present. Of course, these are only hints, and the infer ence may be incorrect. For example, consider the following story:
- (9 ) John wanted to chat with t1ary, but his train left in two days.
Ru le T1 says to infer that there is a goal conflict here, but this is probably not the case. Fortunately, most such situations are detectable with simple consistency checks. For example, in story (9), a consistency check would reveal that chatting usually takes much les s time than is available, so that the pre sence of a conflict is unlikely. He shall not expand on consistency here because they are only of importance in unusual situations.
3.1.2. r1utually exclusive states
Goal s often conflict because they require some states to exist that are inherently inconsis ten t. For example, consider the following stori es :
- John wanted to marry r1ary. He also wanted to marry Sue.
- John wanted to have a regular job. He also wanted to collect unemployment insurance.
~utua l exclusion is similar to Sacerdoti's' "double cross" conflict (Sacerdoti, 1977). A double cross occurs whenever "each of two con junct ion pu rposes denies a precondition of the other" . 11utual exclusion is somewhat different since it applies both to double cross situations and si tuations in which goal states are exclu sive i ndependent of particular plans. In addi tion, states can be either logically or social ly exc lusive . For example, being unemployed and having a job are logically impossible, while being married to two different people at once is exclusive only according to the con ventions of one's society.
3.1.3. Causing~ preservation~
Consider the following stories:
- John wanted to go to the football game, but it was raining outside.
- John wanted to take the night off, but he thought his boss would fire him.
In both cases, the conflict arises because execution of a plan for a goal would cause a character to have the goal of preventing some thing from happening. Thus a goal conflict can occur when the plan for a goal causes a char acter to have a preservation goal. In order to detect these conflicts, a reader must examine the plans likely to be used for each character's goals, and determine if the execution of that plan will invoke a preservation goal for that character.
3.2. Goal Conflict Situations
Goal conflicts give rise to interesting story situations. For example, consider the follow ing situations:
- John wanted to watch the football game, but he had a paper due the next day.
- (a) That night, John watched the football game . John failed Civics.
- (b) John called his teacher and asked if he could have an extension .
- (c) John decided to record the football game on his Betamax.
- (d) John's teacher informed him that the dead line for the paper had been postponed.
These stories are instances of the following types of goal conflict situations:
- Goal abandonment- The character opts for one goal, abandoning the others . This is the case in story (14a) .
- Goal Conflict Resolution- The character can try to resolve the conflict, as in stories (14b) and (14c).
- Spontaneous Goal Conflict Resolution - The conflict may be resolved by some event not initiated by the planner . Such an event occurs in story (14d) .
PN1 uses knowledge about the possible si tuat i ons that might arise from a goal conflict to explai n subsequent events. The following is an instance of a story PAI-1 understands in this manner:
PAll Example:
INPUT: JOHil HAN TED TO \·lATCH THE FOOTBALL GAI·1E. HE HAD A PAPER DUE. THE fJEXT D/\Y. JOHfl \JI\TCHED THE FOOTBALL GAllE. JOHi'l FI\ILED CIVICS.
Question:
INPUT:
tmY DID JOHN FAIL A COURSE?
RESPONSE: BECAUSE HE FAILED TO HAND IN AN ASS IGilr!EilT.
PAf1 uses the presence of a deadline to infer a conflict. Since John pursues one of the goals, PAf·l infers this is a goal abandonment situation and infers the failure of John's other goal.
4. GOAL COilPETITION
Stories usually have more than one character. Goal comoetition arises when the fulfillment of one character's goal will interfere with the fulfillment of the goals of the other character.
Goals can compete with one another for reasons similar to those underlying goal conflict. How ever, the situations that can arise from goal competition are substantially different. Con sider the following stories:
- ( 15) John told f1ary he wanted to watch the foot ball game. f1ary said that she wanted to watch the Bolshoi ballet.
- (a) ilary put on channel 3. John got out the la~mmower.
- (b) John got the old black and white TV out of the attic.
- (c) flary put on channel 3. John punched f1ary in the mouth and put on the ball game.
- (d) John told 11ary he 1~ould take her out to the ballet if she would watch the football game with him.
- (e) i1ary put on channel 3. She found out that the ballet was postponed until later that day.
These stories are instances of the following types of goal competition situations:
- ( l) Independent goal pursuit, in which each character pursues his goals without re gard to the other characters. Story (l5a) contains an instance of this situation.
- Anti-planning, in which a character tries to interfere with another character 's plan. Anti-planning occurs in examples (l5c) and (l5d).
- Easing the competition, a situation where a character tries to undo an underlying cause of the competition. An example of
- this is story (l5b).
- External competition removal, in which the goal competition is removed by some action not taken by one of the competitors, as in story ( l5e).
4.1. Anti-planning
Anti-planning, or the process of countering one's opponent, is the most complex of these story situations. The following forms of ant i planning are distinguished :
- (l) Physical Elimination
- Sabotage
- Persuasion
The plans applicable to (l) and (3) are the general plans for overpowering someone and for persuading someone, which have been discussed elsewhere (see Schank and Abelson (1977) and ~lilensky (1978)). Also, Bruce and Newman (1978) point out that story situations involving several characters often involve deception. In the scheme of things used here, deception is a variation of plans of persuasion in which some of the argument used by the persuader is known to the persuader to be false. These variations are called trick options. For example, a trick option in the BARGAIN plan may be to bargain away some object one does not possess; a trick option in the THREATErl plan is to bluff, or pretend that some object is a weapon that is not (e.g., a toy gun). PAt1 does not currently know about the trick options of any plan, however.
4.1. l. Sabotage
Consider the following story:
- John and Bill both wanted to marry ~~a~y. John began spreading nasty rumors about Bill.
John is using an anti-plan to stymie Bill's goal in order to maintain the possibility of fulfilling his own. This anti-plan is called UflDO-PRECONDITION. It involves pursuing a plan whose goal is to remove a precondition for someone else's plan. For example, in the story above, John employs the UNDO-PRECONDITION anti plan against Bill by trying to get 11ary to dis like Bill, since Bill's plan to get t1ary to marry him probably requires that she be favor ably disposed toward him.
The following is a computer example of PAr1 pro cessing a story involving sabotage:
PAf1 example:
INPUT: JOHN WANTED TO WIN THE STOCKCAR RACE. BILL ALSO WANTED TO lHN THE STOCK CAR RACE. BEFORE THE RACE, JOHN CUT BILL'S INGITION WIRE.
Question:
INPUT: WHY DID JOHN BREAK AN IGNITION WIRE? RESPONSE: BECAUSE HE WAS TRYING TO PREVENT BILL FROM RACING.
Because John's goal and Bill's goal constitute mutua l ly exclusive states, PAM infers the pre sence of goal competition. Since having a vehi cle in good condition is a prerequisite for bei ng in a race, PAM infers that John is using the UNDO-PRECONDITION plan to anti-plan against Bill.
4. 1.2. Avoiding Danger
A special set of anti-planning rules is requir ed for understanding goal competition situations based on generating a preservation goal. A par ti cularly interesting subset of these situations occurs when a character may wish to execute a plan so as to avoid causing someone else to have a goal competitive with his. These are ca lled anticipated preservation ~situations. For example, consider the following stories:
- (17 ) John wanted to get the treasure, but it was guarded by a huge dragon.
- (a) John tiptoed past the beast.
- (b) John threw a rock on the other side of the beast and then got the treasure when the dragon ran over to investigate the dis turbance.
- (c) John poisoned the dragon's lunch.
Each of these stories demonstrates a plan that can be used to execute another plan without generating an anticipated preservation goal. These plans are given the following names:
- (l) AVOID-DETECTION- executing a plan in such a way so as not to be detected by one's opponent. This is the case in story (l7a).
- DISTRACT - turning one's opponent's attention elsewhere. In story (l7b), John executes a plan to divert the dragon's attention, after which John will presumably pursue his plan to get the treasure.
- OVERPOWER- physically pre-empting one's opponent. John tries to reduce the dragon 's physical abilities in story (l7c) so that t he execution of his plan for getting the t reasure does not generate a preservation goal.
The following is a computer example of PAr1 understanding a story involving an anticipated preservation goal:
PI« example :
INPUT: JOHN WANTED TO GET THE TREASURE, BUT IT WAS GUARDED BY A DRAGON . JOHN WALKED OVER TO THE TREASURE QUIETLY .
Question:
INPUT: WHY DID JOHN SNEAK OVER TO THE TREASURE?
RESPONSE: BECAUSE HE WAS TRYING TO GET THE TREASURE AND HE WANTED TO PREVENT THE DRAGON FROM KNOWING WHERE HE WAS.
PAM infers two explanations for John's behavior. First, that he is executing a plan to get near the treasure, and second, that he was also using the AVOID-DETECTION anti-plan to prevent generating opposition from the dragon.
5. CONCLUSION
Stories are typically about interesting dramatic situations . Many such situations are based on interactions between goals. In order to build an understander that can comprehend stories about real world situations, it is necessary to have a theory of plans and goals that reflects the complexity of the real world. PAM uses such a theory to find explanations for characters' behavior. However, such knowledge should also be applicable to a broader range of AI problem areas. For example, problem solving requires a planner to consider situations in which a num ber of goals are present. The theory of goal relationships presented here should be useful for planning for and acting in these situations as well as for understanding them.
@inproceedings{ijcai1979p210,
title = {Understanding Complex Situations},
author = {Wilensky, Robert},
booktitle = {Proceedings of the Sixth International Joint Conference on
Artificial Intelligence, {IJCAI-79}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Bruce G. Buchanan},
year = {1979},
month = {8},
url = {https://ijcai.org/proceedings/1979/210},
}
TY - CPAPER TI - Understanding Complex Situations AU - Wilensky, Robert T2 - Proceedings of the Sixth International Joint Conference on Artificial Intelligence, IJCAI-79 PB - International Joint Conferences on Artificial Intelligence PY - 1979 UR - https://ijcai.org/proceedings/1979/210 ER -Copied
Machine-readable versions of this paper, for tools and AI agents: