---
title: "Jack and Janet in Search of a Theory of Knowledge"
authors:
  - "Eugene Charniak"
doi: ""
url: "https://ijcai.org/proceedings/1973/43"
pdf: "https://ijcai.org/proceedings/1973/0043.pdf"
venue: "Proceedings of the Third International Joint Conference on Artificial Intelligence"
year: 1973
pages: "337-342"
track: "Natural Language; Semantics and Parsing"
keywords:
---

# Jack and Janet in Search of a Theory of Knowledge

Eugene Charniak

*Proceedings of the Third International Joint Conference on Artificial Intelligence*, pages 337-342.

## Abstract

In order to ans wer qu e stions a bou t children's stori es one ne e ds a ~reat de al o f "co mmo n sense" knoviledp;e. A model Is rresented which g ive s a rou gh or ~ anization to this knowledge alo ng with specifications as to how the infor ma tion will be accesse d . This r o ug h mode l is then use d as a basis for ti gh t arguments about narrow Issues (primarily using examples concerning ri gg y banks.) The paper is inten ded as an illustrati o n o f how one mi gh t go about constructin g a theor y of knowledge.

## Full text

## 1 Introduction

Let us consider the prob lem of constructing an abstract model of story comprehension. To determine what the model, or· progr am, has "understood" about what it has r e ad, we will ask it questions. So a t ypica l story might start:

- (1) Janet needed some mo ney. She go t he r pip;gybank ( PB ) and started to shake it. Finally some money came out.

Some t ypic al questions would be :

- (2) Why did Janet ge t the PO?
- (3) Did Janet get the money?
- (4) Why was the PB shaken?

0uestions (2) - (4) are no t answered explicitly in the text. That is, the story did not say "Janet got he r Pn because she ... "The story does no t even cont a in a full implicit answer; one canno t logica lly deduce an answer from the statements in the s t o ry without usin g general knowledp; e about the world such as:

- (5) One can often get money from PGs.
- (6) The hard part of getting money from a PB is getting it out. Once that Is done one can be said to have the money.
- (7) Shaking helps get money out of a PB,

So in order to understand a chil dren's story we need a theory of every day knowledge . This theory woul d have to answer questions 1 ike "What is the knowledge we have?" and "How is it or ganized so we can get at the necessary information when it is needed?" Note that this latter question assumes that we have some specific task or tas ks in mind, in our case answering questions about children's stories.

The rest of this paper divides into t wo parts. In the first part a rough description of a model of children's story comprehension will be presented . In the second section we will assume the model presented in the first and look a t some narrow questions concerning the organization and content of our knowledge about piggy banks.

## 2 A~ Qf Children's ~Comprehension

The mode l presented here is solely concerned with deduction and does no t consider problems of natural language per se. In particular it does not deal with syntax or thos e problems on the boundary between syntax and deduction like disambiguation of word senses and determination of noun phrase r efer ents. (However, my Ph.D. thesis considers the noun phrase problem in some detail.)

So we will assume that as the story comes into the program It is immediately translated into an internal representation which is convenient for doing deduction. The internal representation will be a group of "assertions" each assertion being a predicate on an arbitra ry number of arguments. Putting an assertion into the data base is to "ass e r t" i t. The mode 1 w i 1 1 t r y to "f i 1 1 i n the blanks" of the story on a line by line basis. That is, as it goes along, it will try to make connections between events in the story (usually causal connections) and fill in missing f acts which seem Important such as Janet's now having the money in (1).

## 2.1 Demons~~ Routines

Consider a f act like :

- (8) If "it is (or will be) raining" and if "person P is outside" then "P will get wet"

~le have an intuitive belief that (8) is a fact about "rain", rather than, say, a fact about "outside." Many things happen outside and getting wet is only one of them. On the other hand only a 1 i~ited number of things happen when it rains.

We will embody this belief in our system by associating (8) with "r<tin" so that only when "rain" comes up in the story will we even consider using rule (8). We will say that rain is the ".t.Q.Q.Lc. concept" of (8). To put this another way, when a concept is brought up in a story, the facts associated with it are ".!:!l.ill!..e. available" for use in making deductions. nle will also say that the facts are "..P..!.!.t. l.n" or "assert e . " ) So, if "circus", say, has never come up, the program wil 1 not be able to make deductions using those facts associated only with "circus."

Note however that we are not saying that "rain" has to be mentioned explicitly in the story before we can use (8). It is only necessary that there be a "rain" assertion put into the data base. Other parts of the story may provide facts which cause the pror; ram to assert that it is raining. For example:

- (9) One afternoon Jack was outside playing ball with Bill. Bill looked up and noticed that the sky was getting dark. 11 1 think we should !top" said Bill. 11 ~/e will get 1-1et if we keep playin~."

Here, the sky's getting dark in the afternoon suggests that it is going to rain. If this is put into the data base it will be sufficient to bring in facts associ a ted with "ra rn-.--..-

Also note that a topic concept need not be a single "key word." A fact may not become available to the system until a complex set of relations appears in the data base. A fact may be arbitrarily complex, and in particular may activate other facts depending on the presence or absence of certain relations in the story.

looking Forward, looking Jl.2.d. \'/hen a fact is made available we might not have all the information needed to make use of the fact. Since we are making deductions as we go, if the necessary information comes in after the rule has been asserted we want to make the deduction when the information comes ln. So we might have:

- (10) Jack was outside. It was raining.
- (11) It was raining. Jack was outside.
3. In (10) ther~ is no problem. When we introduce "rain" ~t1e have sufficient information to use (8) and deduce that Jack is going to get wet. But in (11), we only learn that Jack is outside after we have mentioned rain. If we want to use (8) we wi 11 need some v1ay to have our fact 11 look forward" in the story. To do this we vlill break a fact into two parts, a pdttern and a

body (an arbitrary program). We wil 1 execute the body of the fact only when an assertion is in the data base which matches the pattern. (~le wi 11 also say that the assertion "excites" the fact.) In (8) the pattern would be "someone outside·." Then in (11) when we introduce (8) no assertion matches the pattern. But the next line creates a matching assertion, so the fact will be excited. vie will say that a fact is "looking forward" 1-1hen Its topic concept appears before the assertion which matches the pattern. When the assertion which matches the pattern comes first we will say that the fact is "lookjng backward" (as in 10).

We can see how important looking forward is with a few examples.

- (12) "Janet was thinking of getting Jack a ball for his birthday. ~lhen she told Penny, Penny said, 'Don't do that. Jack has a ball.'" Here we interpreted the line "Jack has a ball" as meaning that he did not want another. The common sense knowledge is the fact that in many cases having an X means that one will not want another X. This piece of information would probably be filed under "things to consider when about to get something for somebody else." Naturally it was an earlier 1 ine which mentioned that Janet was thinking of getting Jack a ball.
- (13) "Bill offered to trade his pocket knife for Jack's dog Tip. Jack said 1 1 will ask Janet. Tip is her dog too."' The last line is interpreted as the reason Jack will ask Janet. This requires Information about the relation between trading and ownership.
- (14) "Janet wanted to get some money. She found her piggy bank and started to shake it. She didn't hear any thing." The last line means that there was nothing in the piggybank on the basis of facts about piggybanks.

In each of these cases it is an earlier line which contains the information which is used to assign the interpretation. So in (12) there is nothing inherent in the 1 ine "Jack has a bal 1" which means "don't get him another." If there were, something in the 1 ine would also have to key a check for the following situations:

- (15) Bill and Dick wanted basebal 1. When Jack said "There is Jack. ba 11." to play came by B i 1 1 He has a
- (16) Tom asked his father if he would buy him a ball. "Jack has a ball," said Tom.

- (17) Bill's ball of string was stuck in the tree. He asked Jane how he coulrl get it out. Jane said "You should hit it with somethine;. Here comes Jack. lie has a ba ll."

Those familiar with Planner might notice that our "facts" look quite similar to Planner antecedent theorems, with the exception that our facts can "T ook back" as well as "Took forward." Antecedent theorems are only designed to look forward. I Initially formulated facts as antecedent theorems because I was so impressed with the need to "Took fon·1ard." However, rather than call the facts antecedent theorems, I call them "demons" since it is a shorter name.

Specification ~ Removal of Demons. It should be emphasized that the model does not "learn" the information contained in the demons. This information is put in by the model maker. On the other hand, the demo ns are not specific to the story in the sense that they mention Jack, or "the red ball." Rather, they talk about "a person X" who at one point in the story could be Jack, at another, Bill. \·Je will assume a mechanism for binding some of the vari~bles of the demon ("specifying" the demon) at the time the demon is asserted.

We want demons to be act ive only while they are relevant to the story. A story May start by talking about getting a present for Jack, but ultimately revolve around the games played at his party. We will need some way to remove the "present gett ing" demons when they have outlived their usefulness. (/\n irrelevan t but active demon not only wastes time and space, but can cause us to m i s i n t e r pre t a new 1 i n e . ) As a f i r s t approximation we will assume that a demon is declarej irrelevant after a given numer of lines have gone by.

~Routines. So far we have said that demons are asserted whe n the proper concept has been mentior:Jed. But this imp! ies that there is something attached to the concept name telling us what demons should be put in.

If I'Je look at a particular example, say (13), it is Bill's offer to trade which sets up the context for the rest of the fragment. I will assume that the information to do so is in the form of a program. Such routines, which are available to set up demons, will be called"~ routines."

These base routines will be responsible for more than setting up demons. Suppose we are told that Jack had a ball , and Bill a top. Then Jack traded his ball to Bill for the top. One question we might ask is "VIho now has the top?" Natura lly since questions of "who has what" are important in understanding stories we will want to keep tabs on such information. In this particular case, it must again be the "trade" statement which tells us to switch possession of the objects. Every time a trade occurs we will want to exchange objects, so whenever we see "trade" I'Je execute the "trade" base routine. Of course, the program can't be too simple­ minded, since it mus t also handle "I will trade ... " and perhaps even " ~Ji 11 you trade

... ?"

A good test as to whether a given fact should be part of a base routine or a demon is 1-1hether we need several 1 ines to set it up or whether we can illustrate the fact by presenting a single 1 ine. (Naturally several 1 ines could be made into one by putting "and's" between them, but this is dodging the point. I am only suggesting an intuitive test.) So we saw that "Jack has a ball" was not enou~h by itself to tell us that Jack does not want another ball. Hence this relation is embodied by a demon, not a base routine. On the other hand, often a single line can tell us quite a bit as in "Jack was on second base." This indicates that the base routine for "second base" can often tell us that we are talking about a basebal 1 game.

## 2.2 Bookkeeping~~ Finders

Updating~ BookkeePing. Up to this point we have introduced two parts of the model, demons and base routines. In this section we will introduce the remaining t~1o parts.

/\gain let us consider the situation when Jack had a ball, [l ilT a top, and they traded. ~lhen v1e say that Bill now has the ball, it imp! ies that Jack no longer does. That is to say, we must somehow remove the fact that Jack has the ball from the data base. Actually we don't want to remove it, since we may be asker! "l·lho had the ball before Bill did." Instead, we want to mark the assertion in some way to indicate that lt . has been updated. We will assume that there is a separate section, pretty much independent of the rest of the model, which is responsible for doing such updating. ~le will call this section "bookkeeping."

~Finders. But even deciding that one statement updates another requires special knowledge. Suppose we have:

- (18) Jack was in the house. Sometime later he was at the store.

If we ask "Is Jack in the house?" we want to answer "No, he is at the store." But how is bookkeeping going to figure this out? There is a simple rule which says that (<state> A B) updates (<state> A C) where C is not the same as B. So (AT JACK FARM) would update (AT JACK NEW-YORK) . But in (1 8) we can't simply look for Jack AT <someplace which is not the store>, since he is IN the house. To make things even worse, we could have:

- (19) Jack was in the house. Sometime later he was in the kitchen.

To solve this problem we will need:

- ( 20) To establish that PERSON is not at location LOC.

Find out where PERSON is, call it X. If X = LOC , then theorem is false so return "flo." If X is part of LOC then return

If LOC is part of X, then try to find a different X.

"tlo." Else return "Yes."

In (18) the bookkeeper would try to rrove that Jack is not at the store, and it would succeed by using (20) and the statement tha t Jack is in the house. Bookkeeper would then mark the earlier statement as updated.

Theorems like (20) are called ".lil.tl finders."

Like demons, fact finders have a pattern and a body. A particular fact finder is called when either a demon, base routine or bookkeeping wants to establish a goa l which matches the fact finder's pattern. This is different from demons which are called when we encounter a given fact.

The basic idea behind fact finders is that they are used to establish facts which are comparatively unimportant, so that we do not want to assert them and hence have them in the data base. So in (18) we do not want to assert ".Jack is not in the house" as we l 1 as "Jack is at the store." In the sane way we will have a fact finder which is ab le to derive "<person> knows <fact>" by a sl: ing such questions as "v1a s the <person> there 1·1hen <fact> was mentioned or took place?" Ar,ain, this information is easily derivable, and not all that important, so there ~o/Oul d seem to be no reason to include it explicitly in the data base.

## 3 ~ t!arrow Ouest ions

In section 1 we stated that our theory of knowledge should answer questions 1 ike "hov1 do we access the in format ion" and "what is the information." In this portion of the paper we wi 11 look at tv10 problems, one of each kind. We start with a question of

information access.

## 3.1 Demon-Demon Interaction

In the description of the r.1odel it v1as stated that demons are excited when an assertion enters the data base which matches the demon's pattern. In this section we will present evidence that, r,iven the model of part 1, we must also allow demons to excite other demons. I call this ".f!..e.m.Q.n-demon

interaction."

- A ~ ~ Pfls. Before we can talk about demon-demon interaction we need to establish the need for two particular demons. Suppose we were given:
- (21) Janet needed money. She f',Ot her piggy bank.
- (22) Janet got her PB. "I \'Jant a nic kel" she said.

Were we asked what Janet is going to do with the PB we \·Jould say, "get money from it." The obvious way to handle this is with a demon which is declared relevant when we see a person getting a P~. Naturally this would be done by the PB base routine. This demon would be looking for "<getter> need <money>" (i.e., that would be its pattern). When excited the demon would assert that there is a causal relation between "need money" and "get PB." \'le l'lill call this demon PB-~'EED­ r.lml EY.

Now we nieht claim that "want money" should rut in a demon to look for "get PB" re1ther than vice versa. llovtever, this seems to be less reasonable, since there are ma ny ways of getting money, but only a very

limited number of reasons a person gets a PB.

- A .Q..gmr About Buyjng. Buying things often requires money. So if we saw
- (23) Janet was going to by some candy. She needed some money.
- (24) Janet needed some money. She was going to buy some candy.

we would assert that the reason she needs money is to buy candy. Since the "need money" statement can occur on either side of the "buy" statement, it is clear that we want "buy" to put in a demon which says "If the 'buyer' needs money, it is because of the 'buying'." ~le will call this the BUY-NEED110tlEY der1on.

l'Je do not want "need money" to put in a demon looking for "buy." There are many things one can do with money, bribe a juror, pay rent, take a taxi, tip a bellboy, etc. It would seem more obvious to have the various events state that they (usually) require money, rather than have "need money" look for all of them. Nor does it seem reasonable to claim that all of these events ("bribing", etc.) are really versions of "buying." To express "take a taxi" as "buy some ot the taxi driver's time plus the temporary use of the automobile in orde r to convey onself from one loaction to another" seems somewhat forced. Most of this "definition" comes from knowledge of

economics rather than taxis.

Evidence Qf ~ Phenomenon. consider the following fragment: But now

- (25) Janet was going to buy some candy. She went to get her PB.

!!ere we want to assert that Janet gets her PB because she wants money. But this time there is no previous assertion which says that she needs money. Of course, what is at work here is the fact that buying requires money. But, we have represented this fact as a demon, and so far we have no way for two demons to interact with one another. Now both PBt!EED-r.lotJEY and BUY-NEED-~10NEY wi 1 1 have the same pattern. So we can account for (25) by allowing a demon to excite other demons with the same rattern. Naturally this is what I mean by demon-demon interaction.

A Restriction Qll ~-Demon Interaction. Demon-demon interaction is probably more complex than we have indicated so far. Consi der:

- (2 6) Janet was going to buy some candy. She was a lso go ing to buy some fruit.

In (2G) bot h occurences of " buy" wil 1 activate BUY- ~EED-MOH EY demons. (Though we di d no t comment on this earlier, the idea of specifying demons as mentioned In 2.1 obviously r equires separate copies of a demon t o be able to exist simultaneously.) However, (26) does not imply that Janet really needs money. For al 1 we know she has as muc h as she needs in he r poc ket. If demon-demon Interaction were as simp le as we have made it out to be, the two instances of BUY- NEED- t10N F.Y would join up to produce a "need money" assert ion. So it is no t sufficient for t v1o demons t o be lool: i ng for t he s ame patte rn.

Looking at example (25) we note that one of the demons ~ave a reason why Janet mi ght need money, and the second sug~ este d that need ing money was the caus e of a certain ac ti on. So we have:

~J i 11 buy -- > Need money --> \·li 11 get PB

To put this in everyday terms , in (25) we have both a motive for needing money (buying ), and a result of needing the money (go and get PB). In (2 6) we have two mo tives. The na tural sug£est i on is that demon-demon interaction be restricted to ca se s l'lhere ~tie have both motive and r esult.

How do we recor;nize when we have both mo tive and r esult? As it stands now one demon l ooks prett y much like an y other. We mi ght just try to labe l all demons as "mo tive" or "result" vlith respect t o their pa tt e rn. On the othe r hand it mi ght be pos s i ble t o de rive "motive " and "result" from mo r e ha s i c cons ide rations. At any r a t e , it s eems premature to forma lize suc h concepts a t t h is point. We si mply don't know enough.

Capturine Generalizations . Before mo ving on I shoul d point out that the k in d of a r gument used in this s ection (and the next a l so) is a "capture the p; eneral ization" type a r gume nt commonly foun d in 1 inguistics. We coul d have crea ted a new demon to explain (25). It woul d have s a i d, "if a person gets hi s PB look for hi m planning to buy something ." However, this woul d be missing the generali zation that "motives" and "results" always act this v1ay. So far I have only g iven one example to· support the " demon­ demon int.er action general i z a tion", but in in the ne xt secti on we wil 1 see another.

## 3 · 2 P u t t i n g ~ .in.t..Q .a E.l.z.g,y ll.n.!s.

In this section we will look at a po s s i ble demon associ a ted wi th pi gg y banks and a r gue tha t the deduction it woul d account f o r can be be tter handled by demon-demon int e r action be tween two othe r demons. In eff ec t we wil 1 be trying t o determine , on an ex tremely small scale, what people knov1 .

A ..e.im ill.k Problem. One fact we know about PB's is that they are good places to keep money. This fact seems to come into play in:

- (27) Penny s a id to Janet, "Don't take your money with you to the park. (You will lose it.) Go and get your PB!"
- (28) After Janet helped Ms. Jones with her groceries Ms. Jones gave her a d ime. Jack came along and said, "Come with me to the park, Janet." "OK," said Janet. " But first I am going home to fInd my PB. I do not want to ta ke the money to the park."
- (29) Janet put some money on the sink . Mo ther sai d, "If you leave the money there i t ma y fa 1 1 i n the drain . Let's fin d your PB."

In each case the natura l question is, "Why should Janet get her PB?" Now we mi ght try to construct a "pi ggy bank" demon which responds to some common element in (27) - (29) and then ma ke the necess a ry assertions. A close loo k a t the examples even g ives a start at what such a common eleme nt mi ght be, SiJY "a particular location for the money is negatively evalua ted." ~J e will call this demon PB-BAD-PLACE. The troub le with such a solution woul d be that i t woul d no t account for:

- (30) Janet said, 11 1 am going to put my money away. I will get my PB."
- (31) Janet helped Ms. Jones with her groceries. Ms. Jones gave J anet a dime. Jac k came along and said, "Janet, let's go to the park." "OK," said Janet. "But I want to put my money i n a s afe place. I am go ing to get my PB."

Now there is nothing saying that our demon needs to account for ( 30) and ( 31). llowever, it seems quite obvious that we are using the s ame i nformation in all the examples above. The only d ifference is tha t in (27) - (29) we are expressing the need for a "sa fe place" by making negative comments about another location. If this is a s i ng le fact we would 1 ike a single demon t o express it. The trouble is finding what (27) - (31) have in common.

A l:!.Qn-~ ll.n.!s. Prob I em. In the course of looking at examples I ike (27) (31) I noted examples like:

- (32) Penny s a id to Janet, "Don't take your money with you to the park. Put it on the shelf."
- (33) After Janet helped Ms. Jones with he r groceries Ms. Jones gave her a dime. Jac k came along and said "Come with me t o the park, Janet." "OK ," s a id Janet. "But first I am going to put my money in the house. I do not want to ta ke the money to the par k."

- (34) Janet put some money on the sink. Mother said, "If you leave the money there i t rna y f nll i n the d r a i n . " Janet put the money in a dr c>we r.
- (35) Janet said "I am going to put my money away. I will put it in my toy box."
- (36) Janet helped Ms. Jones with her groceries. Ms. Jones gave Janet a d ime. Jack came along and said "Janet, let's go to the park." "OK," said Janet. "But I want to put my money in a safe place." Then Janet went into' he house and put the money in her room.

These examples exactly mirror (27) - (31), except that (32) - (36) don't mention PB's. Naturally, in these examples the question to ask is "vJhy did Janet put the money in the drawer?", "in the house?", etc.

Such examples tend to indicnte that the problem facin g us is wider that just PB's. vie will name this wider problem the \_"put al·tay" problem. However It is nof- the case that our problem with PB 1 s can be completely reduced to the "put away" problem. So while In the non-piggy bank examples we mention that Janet has or actua 11 y intends to "put" the money some place, in the PB examples all we needed to say was that Janet was going to get the PB. To put this another way, our knowledze of PB's allowed us to Interpret "get PB" as meaning that Janet ~·tas eo i ng to put money into it. However our knowledge of houses or shelves does not allow us to make similar deductions In (32) - (36).

.Illi; .E.ill.-~ ~. banks for the moment, what to (32) - (36) look 1 ike? demon, called the PUT-AWAY activated by 1 ines 1 ike: Ignoring piggy would a solution We wil 1 have some demon, which is

- ( 37) Don't leave the money by the sink.
- (38) I do not want to take my money to the park.
- (39) I will put my money away.

These lines will put In a demon look ing for "put av-tay" and the demon wi 11 assert that the reason for putting the thing away is (37) - (39). Ultimately we will wa nt a theory of why people put things away (i.e., what lines put in the "put away" demon), and how to determine what constitutes a "put awny" location. However, (32) - (36) clearly show that the problem is distinct from the question of what we know about PBs.

.Illi; .!:..i....g,g, .!i.9.nk ~. \·/hat we wi 11 now see is that If we assume the PUT-AWAY demon, all the examples in (27)- (31) fall out easily, plus a few others which we haven't even looked at yet. But first we need to consider a new PB demon entitled PB-MONEY-IN. It is parallel to PB-NEED-MONEY , but while the latter was for recognizing that money was going to be taken out of' he PB, PB-MONEY-IN is for recognizing that money is going to be put in. It says "if you see that the person wants some money to be In the PB then the reason he is getting the PB is to put it in." (Actua lly this theorem is true of a wide - class of containers, but that does not affect the argument at hand.) This demon wil 1 account for examples 1 ike:

- (40) Ms. Jones p;ave Janet a dime. Janet went to get her PB. 11 1 want the money to be in my PB ," she thought .
- (41) Janet got her PB and dropped some money in.
- (42) After Ms Jones gave Janet a d ime, Jack came by and asked Janet if she wanted to go to the park. "OK," said Janet. "I will go home first and get my Pl3." Soon Janet came back and said 11 ~1y money is in the Pl3, 1 e t' s go !"

~-~ Interaction. Now , if we assume demon-demon interaction as discussed In section 3.1, P B -MON EY-1~ plus PUT-AWAY will Interact to solve all the examples from (27) t o (31). Let us see how this will happen.

First note that the restrictions we placed on demon-demon interactions are met here. First both demons have the same pattern, e.g., "money is in PB." (Actually the pattern for PUT-AI·JAY is " ( o-bject > is in <approrriate location>" however <obj ect > will be bound to the money at the time the demon is asserted, and <appropriate location> will match PB when the demon is excited. ) Secondly, we need both a motive and a result before we can "combine" demons. In the case at hand, PUT-AWAY is a motive for having the money in the PB, and "get PB" is a result of intending to put money in the PB.

Saying~. Finally, note that our solution extends to the following ca~e:

- (43) Janet got a dime from Ms. Jones. She said "I am saving my money to buy a bicycle. I am going home to get my PB."

He re we know that Janet is go ing to put the money In the PB because of the "save" statement. However, we immed iately note that we have cases 1 ike:

- (44) Janet got a dime from Ms. Jones. Janet bld her "I am saving my money to buy a bicycle. I am going home to put the money away. (I am going home to put the money in my drawer.)"

tlaturally, (44) indicates that "save" must activate PUT-AWAY. If this Is the case, then (43) is accounted for in exactly the same manner as all the initial examples. I'Jh ile the reader may not be surprised at this resuTt, I am, since initially I thought that the relationship of "save" with pi ggy banks would need a separate PB demon.

## 4 Conclusion

The two halves of this paper stand in contr as t to e ach other. The presentation of the node ! (secti on 2) is ge ne ral (in t heory covering all of chi ! d ren's stories), but vague and full of covert appeals to the reader's intuition. Section 3 on the other hand is na rrow, only tal king about small portions of our knowledge of PBs, but ti r, ht!y re Cl soned (hopefully).

Now by th emselves the conclusions of section 3 a r e not that important. Of course, if we coul d p in down one hund r ed f acts the way we pinned down one in section 3.2 then we woul d have the bep, innings of a theory of knowledge . But I d id not l'·lrite this paper t o tell of one f act about PBs. Rather I view the pape r as an illustra tion of how one mi r, ht go about the ta s k of constructing a theory of know l edge.