Understanding Stories Through Morals and Remindings

UNDERSTANDINS SIORIFS 'IBRCXJGH MJRALS AND REMINDINSS

Michael G. Dyer

Computer Science Dept . 3731 Boelter Hall University of California (UCLA) Los Angeles, CA 90024 U.S.A.

In contrast to IPP and F.RUMP (which as skimners read many stories but built relatively sp:~rse con­ ceptual representations for them) , OORIS (Lehnert, Dyer et al. 1983) was designed to read a few stories in great depth, extracting a detailed con­ ceptual representation involving scripts, plans, goals (Schank and Abelson 1977), settings, affects (Dyer 1983b) interpersonal relationships, and plan­ ning failures organized in terms of abstract thematic structures, called TAUs (Dyer 1981).

This p:~per discusses extensions to BORIS which involve remembering previous episodes in memory during the comprehension of novel stories. 'Ihese theoretical extensions are under design and developnent in a program called MJRRIS (Moral and Reminding Inference System) · MORRIS is intended to read a story in depth, and perform a careful analysis of the appropriateness of character ac­ tions. As a result of this analysis, MJRRIS will extract the JTOral of the story, in terms of abstract planning advice, and use this JTOral as an indexing structure for storage of the story in long-term episodic memory. Whenever a later story is read, which can be analyzed in terms of the same abstract planning advice, MJRRIS will be reminded of the prior story. FurtherJTOre, MJRRIS should be able to express its advice in terms of an appropri­ ate cultural saying, or adage. Unlike IPP, remind­ ings which occur in MJRRIS will be based upon recognizing inappropriate planning situations at an abstract level. 'Ihus, remindings in MORRIS need not be based simply on indexing by roles and

frames.

~. Indexing Stories in ~isodic ~

Often, when people read a story, they are spon­ taneously reminded of previous stories (or personal experiences). Ml.y is this the case? What makes a story memorable and worth telling over and over again (as in certain p:~rables and fables), or worth recalling in the right context?

In this p:~per I argue that one class of remind­ ings occur as the result of a complex analysis of the appropriateness of character plans, actions, or reactions to various goal situations, interpersonal relationships, and other conceptual structures of significance within the story. When a character responds inappropriately or ineffectively, the er­ ror is recognized by the reader, and the conse­ quences of that character's actions are predicted. As a result, the story carries with it a JTOral, or

ABSTRAcr

This p:~per examines the role of lx>th episodic memory and abstract planning knowledge in narrative comprehension. Many narratives are memorable for the abstract planning advice they contain, usually in the form of failures which narrative characters experienced during either planning, or plan execu­ tion. The planning advice, or JTOral, contained within such stories is often expressed in terms of sayings, or cultural adages, which then effectively characterize these stories. 'Ihese same planning failures also provide indexing structures for lx>th the storage and subsequent retrieval from long-term episodic memory of narrative episodes involving similar errors in planning. Consequently, when people encounter similar episodes (either in real life situations, or vicariously in narratives) they experience spontaneous remindings (Schank 1982) of related stories which illustrate the same moral. A theory for extracting the moral of a story, along with remindings of related episodes, is currently begin designed and developed in order to extend the theoretical and practical scope of the OORIS (Dyer 1983) story understanding program.

8. Background

Previous story understanding systems developed at Yale displayed their comprehension by performing tasks of translation, summarization (e.g. F.RUMP (DeJong 1979)) or question-answering (e.g. QUALM (Lehnert 1978)). However, after reading a story and performing its respective task, these programs threw away the results of their work. Consequent­ ly, previous stories were never involved in subse­ quent canprehension processes. IPP (Lebowitz 1980) was one .exception, since it retained a summary of each story, which it then used in forming generali­ zations. For example, if IPP read three stories of bombings in Ireland, and in each case carried out by the IRA, then IPP formed a generalization that banbings in Ireland are done by the IRA. '!he next time IPP read alx>ut banbings in Ireland, it used this generalization to predict IRA responsibility.

While IPP represented an i.rrq;x>rtant advance over other story understanders which failed to remember what they had read, IPP was limited in lx>th the depth of its understanding and the way it indexed stories read. As a skinmer, IPP mainly filled roles in frame-like structures (Minsky 1975). As a result, IPP could only index stories by means of similarities and differences in role bindings.

implicit advice in planning, which the reader may be able to employ in real-life for future planning in related situations. The most generally applica­ ble types of advice becane solidified in the ad­ ages, or sayings of our culture. Such adages, for instance, give us abstract advice concerning:

• timing of plans ("A stitch in time saves nine.")
• plans for satisfying recurring goals ("Kill the goose which lays golden eggs.")
• assessing risks in planning ("The cure is worse than the disease.")
• coordinating planning with others ("Too many cooks spoil the broth.")
• counterplanning and vulnerability ("Throw­ ing stones when you live in a glass house." "Cutting off your nose to spite your face").
• affect reactions during plan/goal failure ("Crying before the milk is even spilt.")

In BORIS, such planning advice is organized by TAUs (Thematic Abstraction Units) which represent situations in which planning errors brought about goal failures. See (Dyer, 1983).

.c. Remindings ~ Comprehension

By way of illustration, consider the following episode (which actually happened to the author):

GARAGE TRIP: I was taking a friend to a garage to pick up his car. He started talk­ ing about how they had pranised to have it ready, but he hadn't called them to make sure. He became convinced that the car wouldn't be ready, and how annoying that would be, and he became more and more upset. When we arrived, his car was all fixed.

At the IX'int that my friend began getting really upset, I told him to quit bleeding before he'd even been cut. I was also spontaneously reminded of a humorous story I had not thought of in years (and whose author or source I can not recall) · Of course I now told him the story as we drove to the garage. 'I have tried to recapture below the flavor of this story as it was first told to me:

OOT OF GAS: A man was driving along a lonely road late at night when his car ran out of gas. He remembered seeing a farm house a mile back, and so started walking toward it. As he walked along he thought: "It's pretty late. If saneone were to awaken me at this hour, I might be pretty annoyed." He kept on thinking along these lines: "The farmer will have to get dressed and siphon gas out of his tractor for me. He may find that very incon­ venient ··· " As he approached the farmhouse he became more and more annoyed.

The farmer was awakened by loud knocking at his door. \'41en he went cbwn to answer it, there stood a man who barked: ~o needs your stupid gas anyway!" and then stanped off into the night.

Clearly, both the actual GAP.JIGE TRIP episode and the OOT OF GAS humorous story involve car troubles. However, there are numerous other instances of car breakdowns that I could have recalled, and yet they did not come to mind at that time. The key to my reminding was the moral, or planning advice in OOT OF GAS. This advice was awropriate to GARAGE TRIP and is expressed in our culture by such adages as "Crying before the milk is even spilt" or "Bleeding before one's been cut". B.lt what was the inechanism which allowed me to extract both a moral and a rem­ inding fran the GARAGE TRIP episode?

At the level of abstract planning, in both cases sane individual X became emotionally upset at a si­ tuation S which involved anticipating a goal failure before X knew for sure that S had occurred. Thus, the moral of each story might be expressed in planning terms as follows:

TAU-EMOT-ANTICIPATE: When anticipating a negative situation, don't act as if the failure has occurred, since this wastes ener­ gy that could be better used in planning ways of recovering from the failure, if indeed one has occurred. Furthermore, getting upset may make things worse, by interfering with plan­ ning about what to do next.

In OOT OF GAS, the driver was so upset that he rejected help before even requesting it.

This planning advice is rather sophisticated, involving interaction between plan failures, the appropriateness of anticipating failure in plan­ ning, and the effect of emotional reactions in the face of planning failures.

Upon realizing that my friend was wasting ~nergy and suffering emotionally in anticipation of ~~ unproven failure, I had indexed the situation in terms of the appropriate abstract planning theme. At this point, both the adage and narrative episode which shared the most content features (i.e. a story illustrating TAU-EMOT-ANTICIPATE which also involved a broken car) came to mind. So, not only was I able to advise my friend of the inappropri­ ateness of his behavior in terms of its overall ef­ fectiveness, but I could also illustrate my advice with a vicarious episode from my own experience.

.D. ln-.Del2tb Understanding .tbr.u .fl.an AnalYsis

This level of story understanding is one of the deepest IX'SSible (for other thematic levels, see (Lehnert 1982), (Schank 1982) (Wilensky 1981)). Summarization, translation and question answering tasks are not guaranteed to reveal this level of canprehension. For instance we could ask the fol­ lowing questions regarding OOT OF GAS:

• Q: What was wrong with the driver's car?
• A: It was out of gas.

• Q: Why did the driver go to the farmer's house?
• A: To get help.
• Q: What woke up the farmer?
• A: The driver knocked on his door ·

and still miss the point of the story. The best way to reveal this level of understanding is through spontaneous remindings of episodes with similar themes, along with the generation of advice in the form of a moral, saying, or adage. This deep level of analysis requires sophisticated knowledge of plans, including:

• the goals they normally serve
• a history of successes and failures for each planning situation
• their enablement conditions, along with timing, effort, etc.

Consider how M)RRIS will work for the following, much simpler story, which involves a planning error in terms of cost:

FACE LIFI': Bill was a poor student. One day he discovered that he had a single acne pim­ ple on his dlin. He decided to go see a plastic surgeon and have his face lifted.

Here Bill chose an overly costly plan to achieve a goal which could be achieved by equally effec­ tive, yet simpler and less troublesome measures. One adage characterizing this situation is: "Kil­ ling a fly with an elephant gun."

The moral: Don't dlose an expensive plan (in time, rroney, physical suffering, etc.) when a cheaper, equally effective plan is available.

In this case, plastic surgery is used to elim­ inate serious facial conditions and involves great discanfort. Since the goal is minor, this plan is inawropriate. As foORRIS notices that an expected, more reasonable, plan was not chosen, various plan­ ning metrics (Dyer l983a) are analyzed, including the cost of the plan, the goal it normally serves, etc. Associated with various planning errors are larger thematic structures (TAUs), which generate more expectations concerning plan outcanes, and carry planning advice. Once a TAU associated with this plan error is recognized, M)RRIS will express the TAU in terms of the "elephant gun" adage, and be reminded of any other story which involves the

same kind of planning "overkill".

M)RRIS is currently under design at UClA, using the insights gained from experience with BORIS at Yale. Many representation and process issues remain which are best resolved through experimenta­ tion with initial implementations. MORRIS will be given an episodic memory of different stories, each indexed in terms of distinct TAUs, along with vari­ ous planning metrics. ~enever a plan, goal, af­ fect, or interpersonal expectation is violated, as­ sociated TAUs will be activated, and these struc­ tures will determine how the story is to be in­ dexed. Research is also underway for building a

generator to produce appropriate adages, and thus act as an advice giver in various planning si tua­ tions.

.f;. Suumary

One important type of story is memorable for the abstract planning advice it contains. This abstract planning advice serves as a major indexing structure for both storing and retrieving stories fran long-term episodic memory. The most general forms of abstract planning advice are often ex­ pressed in terms of carunon sayings, or adages, which refer to the moral, or point, of the story. When similar planning errors occur, the same plan­ ning analysis leads to both recall of associated adages, and to spontaneous remindings of stories which are related at the same abstract thematic level. Thus planning information becomes available for both advice-giving, and the anticipation, avoidance, or recovery of future planning errors.

REFERENCFS

DeJong II, G. F. Skimming Stories in Real Time: An Experiment in Integrated Understanding. ·-TeCh Rep 158, CS Dept. Yale Ph.D. 1979.

Dyer, M. G. The Role of TAUs in Narratives. Proceedings of Third Annual Conference ~ Cognitive Science Society, Berkeley, CA 1981.

Dyer, M. G. In-Depth Understanding: .~ Computer Model of Integrated Processing for Narrative canprehension. MIT Press. Cambridge, MA. l983a.

Dyer, M. G. The Role of Affect in Narratives. Cognitive Science, l983b (in press).

Lebowitz, M. Generalization and Memory ~ an In­ tegrated Understanding System. Tech Rep 186, CS Dept. Yale Ph.D. 1980.

Lehnert, w. G. The Process of Question Answering. LEA, Hillsdale, ID 1978. -

Lehnert, w. G. Plot Units: A Narrative Surranariza­ tion Strategy. In Lehnert & Ringle (eds), Stra­ ~ for Natural Language Processing, LEA, Hills­ dale, ID 1982.

Lehnert, W., s. Harley. Narratives. 1983. Dyer, M., Johnson, P., Yang, c. and IDRIS - An In-Depth Understander of Artificial Intelligence, Vol 20, No l,

Minsky, M. A Framework for Representing Knowledge. In P. Winston (ed), The ?sycholoqy ~Computer Vi­ sion, McGraw-Hill, NY 1975.

Schank, R. Dynamic Memory: ~ Theory ~ Reminding and Learning in canputers and People. Cambridge Univ. Press, MA 1982.

Schank, R. and Abelson, R. Scripts, Plans, Goals, and Understanding. LEA, Hillsdale, ID 1977.

Wilensky, R. Meta-Planning: Representing and Us­ ing Knowledge About Planning in Problem Solving and Natural Language Understanding. Cognitive Science. Vol. 5, 1981.