Cancelled Due To Lack Of Interest

1. Introduction

Not everythirg we read is interestirg. Even within a basically interestirg piece of text, some concepts will seem much more important than others. Some stories will carry so little interest that they can be ignored altogether. A key issue in studyirg the variable nature of interest is to understand \ohy certain pieces of text are not interestirg. After all, a priori, we might expect anythirg included in a story to be ~rth our full attention.

As a simple example of varyirg interest, consider story Sl.

Unidentified gunmen Monday ambushed a pro-Iraqi politician, riddlirg him and his bod~uard with bullets, in the latest of the region's almost daily political assassinations.

Police sources said Mussa S'lueib, a tx>p official of the Lebanese chapter of Iraq's rulirg Baath party, was drivirg from Beirut airport When his car was overtaken by armed men and sprayed with gunfire in a gargland-style hit.

Shueib and his bod~uard died on the spot.

In Sl, the death of the bod~uard ~uld usually not be considered tD be as interestirg as the killirg of S'lueib (it is overshadowed). Neither is the method of the killirg as a shootirg too interestirg, as it can be reconstructed from general knowledge of terrorisn. Several other points are simply irrelevant. H3virg dispensed with the non-interestirg, we are left will the interestirg roncept of an important pro-Iraqi politician belrg killed in Lebanon.

It is important tD recognize that interest depends upon the current state of memory. - Interest is highly idiosyncratic from reader to reader and from time to time. Concepts interestirg to one reader at a given time may not be tD other readers or at different times. For exCI!Iple, a read.er who knows Shueib's bod~uard personally ~uld find the pieces of Sl about him quite interestirg.

A computer Ll'lderstandirg system will require lorg-term memory, such as that of IPP (the Integrated Partial Parser) [2, 3], to be able to make use of interest. fbwever, with this memory, it can use interest to help decide what to remember and to guide the understandirg process. It was in the context of IPP that the ideas introduced here were developed. In Section 3 I will describe the uses that interest can be put to.

Generally, a testing process that concentrates on recogni zirg non-interest will be quite effective. For exCI!Iple, an algori tl1n that recognizes an instance of non-interest can immediately reduce or terminate its processirg effort. The alternative, however, involves assllllirg that an item can be ignored only when it does not satisfy positive criteria for interest, a test that may produce an answer too late to be useful.

2. The Nature of (non) Interest

Interest might seem to be an arbitrary and hard­ to-define idea. fbwever, some confusion is simply due to the fact that there are several distinct ways that a concept can be (or can fail to be) interestirg. Figure 1 lists three ways a concept can fail to be interestirg.

Figure 1: Grounds for non-interest

1. Concepts not causally connected to the main point of the story.
2. Concepts that can be reconstructed.
3. Concepts that are overshadowed.

Lack of causal connection (see [ 4]) and reconstructibility (see [1]) explain a large number of situations where a piece of a text is not interestirg. These are in some sense "logical" reasons for findirg a concept non-interesting.

Here I will just present an example of each case.

52 - UPI, 21 June 1980, El Salvador

A right-wing terrorist group called the Squadron of ~ath killed 10 men Saturday incliJdil'l1 a labor leader and three others shot to death as they ate breakfast in a restaurant, pol ice said.

In story 52, the fact that the victims of the terrorist shooting were eating breakfast is oot connected to the main causal chain of the story, and hence oot interestillJ. 'lhis part of the story is oot important to our overall l.l'lderstanding, and will be quickly forgotten.

53 provides an illustration of reconstructible concepts.

53 - UPI, 20 June 80, Guatemala

The terrorists who kidnapped a Nestle Cbrp executive said Friday he will be released only if the SWiss food firm canes up with an undisclosed ransom and pays for the publication of a terrorist manifesto.

In this story, the terrorists aski03 for ransom is quite typical of kidnappings, and so can be easily reconstructed from the simple knowledge that there was a kidnapping. As with the irrelevant part of 52, the reconstructibility of the ransom demand of 53 makes it not interesting. The demarx:J for ransom will oot be long remembered, nor need it be processed in detail.

The third idea in Figure 1, overshadowing, lacks the obvious rationale of irrelevance and reconstructibili ty. 'lbnetheless it plays an important role in interest evaluation. Consider the following story.

54 - UPI, 29 July 1980, Madrid

Leftist gunmen crnbushed an army general in the center of Madrid today ~l.l'lding him but killin;~ a soldier in his escort, police said. Unofficial reports said one of the general's aides and his chauffeur also were ~l.l'lded in the early morning attack.

Normally the killing of the unnaned soldier ~uld seem less interesting than the ~l.l'lding of the general and the various injuries less interesting still. lbwever, this cannot be explained by either reconstructibility or non-causality. 'lhe death of the soldier arx:J the ~unding of the aides are certainly causal results of the attack. Furthermore, the results surely cannot be reconstructed from memory. 5::> why is it that the death and other injuries are so much less interesting?

Stated simply, the attack on the general overshadows the other events. It is oot that a soldier beil'l1 killed is inherently non-interesting but that the attack on the general . is roore interestillJ. Even too~.gh an event might be interesting by itself, most processing will devoted to an overshadowing event.

A concept can have sufficient interest to overshadow for at least t~ distinct reasons. It can be either important or unique. Fbr example, the event, "Terrorists shot the President," is interesting because the victim of the shooting is important. This is contrasted with an event. like, "Three clowns fired soots into a group of spectators," where nothing in the event is very important, but it is a relatively unique event.

Imp:>rtance simply represents a concept for which the reader has a complex structure in memory. 'lhis is extremely idiosyncratic arx:J dependent up:>n the current state of memory. Imp:>rtant concepts for a given person might be his best friend, the President of the United States, arx:J the status of his favorite football tean.

Uniqueness, in contrast, represents a concept that is different from others in memory. It can be used to explain the interest of a shooting by a child, an explosion aboard a space ship, or the US rockey team's upset of the Russians. A key problem to be stooled involves figuring out just oow different an event must be before it seems unique (after all, in some sense all events are unique).

The lack of either imp:>rtance or uniqueness will quite likely result in a concept being overshadowed. It is the specific analysis of just what determines imp:>rtance and uniqueness that will allow interest to be fruitfully applied to understanding.

3. The USes of (non) Interest in Understanding

Interest has effects on both long-term memory for stories read arx:J the parsing process. We have already seen that interest interrelates with memory. Alroost by definition, the parts of a story that are interesting are toose that are remembered. This is an imp:>rtant controls on the accLII\ulation of information in memory. Particularly imp:>rtant is the ability to anit material that can be reconstructed fran memory, thus drastically reducing the quantity of information that must be stored. (See [1, 2] for more oh the storage of redundant information arx:J the reconstruction process.)

'lhe analysis of interest described in this paper does make one important addition to the relation of interest arx:J memory. The idea of overshadowil'l1 will have an imp:>rtant affect on what is remembered. As I will mention below, resources are usually not expended on detailed processing of the less interesting aspects of a story. This will normally result in such items being lost and rot saved in memory. While this result is rot clearly desirable for a computer memory system, it may be necessary for the same reason it occurs in hllllan understarrling to allow the corresporrling increase in efficiency of processing.

So we see that the basic effect of interest on memory is a simple one - interesting items are remE!IOOered, ron-interesting ones are rot. lbwever, the effects on direct processing are more subtle. I will mention here some of the possibilities.

'Ihe primary role of interest in parsing is tD help control processing arrl allow the optimal allocation of resources. This has been discussed tD some degree in [6] arrl [5]. Here I will discuss one very powerful way interest can be used in an understarrling system designed tD process large nunbers of texts.

The simplest use of interest in understarrling is to make a decision tD stop processing a sentence (or a story) that does not appear to be interesting. There is an obvious advantage in doing this. Stopping the processing of a sentence or story clearly gives the maximllll possible gain in efficiency - processing time is redu:ed tD zero for the rest of the i tern! Ard this is clearly sufficient, since as was roted before, low interest i terns usually need rot be remembered at all. This use of interest will give computer understarrlers an ability. that people clearly have - the ability to stop reading something. Informally, a program employing the rules that I will describe below will have the ability to become bored.

'Ihe basic idea is simply to track interest at both a sentence arrl a story, beg inning at 1.11certain arrl continuing until a judgement can be made as to whether the item is or is rot interesting. This is done using the rules described in this paper. If the conclusion is drawn that a story or sentence is not going to be interestill3, then processing of that item can be stopped.

It w::>uld seem at first glance as if this plan should rot w::>rk at all. 'tobne of the rules that I have described differentiate between elements of a story that appear at the beginning and those that appear later. Thus it w::>uld seem possible that the i terns that make a sentence or story significant could occur at very end. U;ing the rule described in the last paragralit, these i terns w::>uld never be processed~ Fbrtunately, in practice the key items do usually occur early on, preslJIIably because authors know that rules such as the one above will be applied. As an exanple of this, we w::>uld be mu:h more likely to firrl a story like, ·The President was shot · rather than, ·rn an official limousine today, lohile on a motorcade throu;~h Dallas, a gunnan in the Book Depository, wh.> was later appreherrled by police, shot the President.·

Out interest heuristics will sometimes fail if the key items are idiosyncratic - not those the author w::>uld consider significant. Fbr exanple, in a story about a building takeover where there were many hostages, the identity of the ·1ow level· people taken hostage w::>uld not be found early in the story, but it might make the story significant to a reader ..no knew one of these people. lbwever, that is exactly the type of situation where a hllllan reader w::>uld miss the key information. Some sacrifices must be made to achieve adequate performance.

4. O:>nclusion

In this paper I have just tou:hed the surface of the problem of interest. lbwever, it is an important problem that justifies further effort. In particular, an understarrling of the purpose arrl mechanisn behirrl overshadowing will allow better control of natural language understarrling prograns.

1. ~lodner, J. L. Retrieval and organizational strategies in conceptual memory: A computer model. Tech. Rept. 187, Yale University Department of Computer Science, 1980.
2. Lebowitz, M. Generalization arrl memory in an integrated understarrling system. Tech. Rept. 186, Yale University Department of Computer Science, 1980. PhD Thesis
3. I.ebowi tz, M. The nature of generalization in understarrling. Proceedin:JS of the Seventh International Joint Cbnference on Artificial Intelligence, International JOint O:>nference on Artificial Intelligence, Vancouver, canada, 1981.
4. Schank, R. c. The stru:ture of episodes in memory. In D. Bobrow arrl A. Cbllins, e:t., ~resentation and UnderstarrliR;J: Stl.dies in CoJnitlve SCieriCe, Academic Press, New York-;-1975.
5. Schank, R. c. · Interestingness: Cbntroll ing inference.· Artificial Intelligence Jl· 3 (1979), 273 - 297.
6. Schan~, R. c., Lebowitz, M., arrl Birnballll, L. ·An integrated understarrler .· 1merican JOurnal of Computational ~inguistics ~· 1 (1980), 13 - 3o. -