Towards a Winograd/Flores Semantics PETER MOTT School of Computer Studies and Department of Philosophy, University of Leeds Abstract. A basic theme of Winograd and Flores (1986) is that the principal function of language is to co-ordinate social activity. It is, they claim, from this function that meaning itself arises. They criticise approaches that try to understand meaning through the mechanisms of reference, the Rationalist Tradition as they call it. To seek to ground meaning in social practice is not new, but the approach is presently attractive because of difficulties encountered with the notion of reference. Without taking a view on whether these are insuperable, the present paper accepts Winograd and Flores' challenge and attempts to lay aside reference and to base a conception of meaning directly in terms of co-ordination and consensus within a linguistic community. Key words. Meaning, reference, disjunction problem, situation theory, synonymy, classification, causal theory of reference, co-ordination.
1. Introduction A basic theme of Winograd and Flores (1986) is that the principal function of language is to co-ordinate social activity: "We use language in human activities and our use of linguistic forms is shaped by our need for effective co-ordination of action with others" (p. 62) 1. Winograd (1985) also makes the same point: "Language is first and foremost a means of doing something- of securing co-operative action and consensus." It is, they claim, from this power of language to organise and co-ordinate other social activities that meaning itself arises. From their perspective to understand meaning is primarily to understand the mechanisms by which consensus is established and maintained. They criticise approaches that try to explain how language links up with an antecedently given ready-made w o r l d - approaches that try to understand meaning through the mechanisms of reference. This Rationalist Tradition (as they call it) they declare to be bankrupt. To seek to ground meaning in social practice is of course not new, for example Wittgenstein (1953) obviously does so. But the approach is presently attractive because of difficulties encountered with the notion of reference. These can be seen, for example, in the problems Fodor (1987, 1990) has had in trying to formulate a satisfactory version of his causal theory in response to the Disjunction Problem (Fodor, 1984; Sterelny, 1990) and in Putnam's arguments against the availability of a notion of reference at all (1981, 1983). Without taking a view on whether these difficulties are insuperable, the present paper accepts the Winograd/Flores invitation. It attempts to lay aside reference and instead to base a conception of meaning directly in terms of co-ordination and consensus within a linguistic community. Suppose we wanted to design a robot, Asimov, that could assist people in a Minds and Machines 5: 69-87, 1995. © 1995 Kluwer Academic Publishers. Printed in the Netherlands.
70
PETER MOTT
workshop- it could be addressed in simple ways so, for example, when George says "Get me a hammer", Asimov gets him a hammer. The question is what we really need for this to be possible. Well, Asimov must be able to classify some noises as utterances of the word "hammer", and some objects as hammers. So Asimov must be equipped with what I call classifiers. Asimov's classifiers must regularly classify the same things as hammers and the same things as utterances of "hammer" as George's classifiers do. Such regularity makes us think of classifiers as mechanisms, at least within Asimov. The key point is this: the classifiers in Asimov only have the task of classifying in the same way as the corresponding classifiers in George. It is required that whatever George calls "hammer", Asimov calls "hammer" as well. It is not required that it is in fact hammers that are so-called. This point is important enough to bear repetition. To introduce Asimov into George's workshop we require only that whatever it is that George means by "hammer", Asimov also means by "hammer". The only relation we need to describe a well-designed Asimov is a means-the-same-by relation between agents and words, not the standard reference relation between agents, words and things. We can seek to get at this relation by an analysis of classifiers. This analysis is done in the next section. Because George must use "hammer" (not, for example, "chisel") if he wants Asimov to get a hammer it follows that there must be some sort of link or connection between Asimov's "hammer" - classifier and his hammer-classifier. It may be very complicated and indirect, but some sort of link between the two classifiers there must be. So having got the idea of a classifier we will have to find a sense in which classifiers for words can be linked to classifiers for things. Links are introduced in Section 3. In Section 4 we show how a means-the-same-by relation can be constructed from a system of classifiers and links. This completes the chain: Winograd/Flores ground meaning in co-ordinated action, the means-the-same-by relation secures co-ordination, a system of classifiers and links secures the means-the-same-by relation and classifiers underlie the whole system. Our project diverges from the causal-informational one. For suppose that a causal-informational theorist were to enquire when George and Asimov mean the same by "hammer". She might set out a programme like this: (1) George and Asimov mean the same by "hammer" if and only if there is something that they both mean by "hammer". (2) There is something that they both mean by "hammer" if George refers to hammers by "hammer" and Asimov refers to hammers by "hammer". (3) So explain the conditions under which George refers to hammers by "hammer" and the means-the-same-by relation comes for free (and there is nothing very interesting about it). We, of course, do not take this route. To be sure we do hold that (1) is true. George and Asimov can only mean the same by "hammer" if there is something that they mean by hammer. As theorists wishing to understand people talking
TOWARDS A WINOGRAD/FLORES SEMANTICS
71
together we need a means-the-same-by relation, but if the present proposal is correct, we don't need a prior reference relation to get that means-the-same-by relation. The theorist is thus freed from the task of trying to explain the connection between words and things. Does this alternative approach solve any of the problems that face the causalinformational theorist? Section 5 argues that it does. We are freed from the travails of the Disjunction Problem. Winograd/Flores criticise a "naive view of language" that sees the world as already divided up into categories and the task of language to attach labels to those categories. They say instead that the "meaning of words and sentences is not ultimately definable in terms of an objective external world" (61). In the detail of our theory we still talk about types of objects, and classifiers classifying types of objects (or sometimes properties of objects, in this paper we draw no distinction between types and properties). Does this mean that we are trying to define the meaning of words in terms of objective categories, and hence are actually contradicting the very spirit of the Winograd/Flores view of language? A last section considers this criticism. An Appendix gives a precise Situation Theoretic model of our theory. The reader can turn to this if she wants the answer to a question of detail that the main body of the text does not supply.
2. Classifiers On car-parks in the Yorkshire Dales there are machines which if presented with a £1-coin will issue a ticket, but if presented with anything else won't. These are classifiers of £1-coins, and we use them as a running example, refining an account of a classifier through a number of definitions. We start by saying why a particular machine M is a £1-ctassifier: (DO) M, on a Dales car-park, is a £1-classifier because if M is presented there with a £1-coin then M will issue a ticket and if presented with anything else then M will return it and not issue a ticket. However robustly M is built it is going to have operating limits which if violated will stop it functioning properly. F o r example it would not work on the Moon because there is not enough gravity, or on the bottom of the ocean because it would be squashed, perhaps it would not work in Arizona because it's too hot or Greenland because it's too cold, and certainly it would cease to work and have to be replaced if fraudulent £1-coins become common enough. The upshot is simple but fundamental: M is not a £1-classifier simplicter, but only within a class of situations. So we try as a first step of generalisation: (D1) Within a class of situations S. M is a P-classifier iff Vs E S, if M is presented
72
PETER MOTT
with a P-object in s, then M will say "yes", and if M is presented with a non-P object in s then M will say " n o " . As it stands this is not adequate because very few classifiers will say either " y e s " or " n o " because very few talk. This may seem a bit nit-picking, even facetious, for of course what we mean is that saying "yes" is a generic sort of response implemented in the car-park machine by accepting the coin and delivering a ticket. Likewise saying " n o " is a generic sort of response implemented in that case by not issuing the ticket and returning the coin. Nonetheless, this means that in fully specifying a classifier we have to say what are to count as its verdicts "yes" and " n o " : (D2) Within a class of situations S. M is a P-classifier with verdicts v iff Vs E S, if M is presented with a P-object in s, then M will respond with Vr, and if M is presented with a non-P object in s then M will respond with vs. A classifier must also have a m e t h o d o f p r e s e n t a t i o n - what has to be done before the classifier "notices" an object in question. For as Fodor (1987) observed with some discomfort, a cow-classifier will not deliver its "yes"-verdict to all cows but only those that come in range of it (The problem was that you cannot say "cow" means cow for X if cows cause X to token " c o w " since innumerable cows will have no such effect, Ancient Egyptian cows, cows not yet born etc.) If x is presented to a classifier in accordance with its m e t h o d of presentation tr, then x has thereby a property; the property of being presented to the classifier in accordance with that method of presentation o-. We can call this property a presentation property and say that x is a tr-object. This gives our final definition of a classifier: (DC) Within a class of situations S. M is a P-classifier of o'-objects with verdicts v iff Vs @ S, if x is a g-object and a P-object in s then M will respond with uv, and if x is a o--object and a non-P object in s then M will respond with vN .
We can say for short that M is a (P, o-, v)-classifier in situations S. This defnition is derived systematically using the tools of Situation T h e o r y in the Appendix, its complexity derives largely from what is (or seemed to be) necessary to get a formal version to "work". We show there that it has the expected properties and that we are not relying on any motion of causality, for Putnam's criticisms of reference bear against causality as well (see chapter 12 of his 1983). AN EXAMPLE OF A CLASSIFIER G o r m a n and Sejnowski (1988) constructed a connectionist network that could distinguish mines from rocks on the seabed. More exactly they took recordings of sonar echoes of mines, and of rocks, digitised them using a frequency analyser,
TOWARDS A WINOGRAD/FLORES
SEMANTICS
73
and trained their network on the resulting vectors of frequencies. The network then performed well on other similarly derived echoes. This network is a classifier. So what does it classify? Bechtel and Abrahamsen (1991) suggest that because the network's inputs were encodings of echoes from real mines and rocks the "network learned to discriminate rocks from mines" (p. 128). Churchland (1989) makes a more modest claim saying it learns to "reliably identify mine and rock echoes from outside its training set" (p. 204). The reader should check his or here intuitions: are Bechtel and Abrahamsen overstating the case when they say the detector was a mine detector, is Churchland's description more accurate? We must specify a class of situation S, a presentation property o-, and a pair of verdicts v. The situations involve ships at sea with echo-sounders, recording ebjects x detected on the seabed and sending some of these recordings back to a laboratory where, in the fullness of time, they are digitised and the resulting vectors typed into a computer M running the network. The presentation property belongs to those objects which are detected by the echo-sounder and whose recordings are sent back to the laboratory etc. The verdicts are the computer printing out " M i n e " or " R o c k " on its display. We are told that the system works, so that we know that in any situation s of the described class (C1) (C2)
if x is a mine & x is presented to M then M prints "Mine". if x is not a mine & x is presented to M then M prints " R o c k " .
Because M satisfies these two constraints it is mine-classifier according to (DC), within the class of situations and relative to the verdicts and method of presentation. Bechtel and Abrahamsen are thus exactly right. Of course the system is not in any way a real-time mine-detector which is the sort of detector that people in ships want. But that is a separate issue altogether. However, by changing the class of situations, we can do it Churchland's way too. The situations consist now of someone in the laboratory typing in vectors of numbers x. We suppose that in all these situations we have: (C3) (C4)
if x is a mine-vector & x is typed into M then M prints "Mine". if x is not a mine-vector & x is typed into M then M prints " R o c k " .
For this reason M is a mine-vector-classifier within the situations in the laboratory. Thus both views are correct but relative to different classes of situation, but to the question whether the network is really a mine-vector classifier or a mine-classifier there is surely no answer. 2
3. Systems of Classifiers and Links T h e r e are word-classifiers for recognising words and thing-classifiers for recognising things. These may be linked. When a word-classifier is linked to a thing-
74
PETER MOTT
classifier we have in place a system of classifiers and links for that word. When two individuals s h a r e - in a sense to be described in the next s e c t i o n - such a system, then they are in a position to use the word to co-ordinate their behaviour. If they do so the word thereby acquires meaning. In this section we say what linked classifiers are. We know that are links between word-classifiers and thing-classifiers because, in pathological cases, we can see they are broken. Luria (1972) recounts the extraordinary and tragic case of Zasetsky, a man wounded in the temporal lobe during the Second World War who then spent the next 25 years of his life laboriously writing and re-writing an account of what his affliction was like (which Luria edited and organised). In general terms, says Luria, Zasetsky lost the capacity for "synthesis and organisation of complex stimuli." But it is much more specific than that: words came apart from their meanings. Zasetsky recounts it himself: " I ' d try and clamp the words to the idea as much as I could. But what torture it was" (p. 74) and " . . . there always seems to be a gap between a word and its meaning. These two are always disconnected and I have to yoke them together somehow. But I can't keep them yoked together for any length of time; they come loose and just vanish into thin air." (p. 92). Sometimes things are familiar but he cannot name them, other times given the name he cannot connect it up with the thing. The system of links connecting word-classifiers to thingclassifiers had been all but destroyed. Zasetsky's cognitive deficit was extreme and general. However it now appears that the links between thing-classifiers are much more specific and localised than his case might suggest, for you can lose some of the connections between words and things while keeping others. For example "colour anomia" is a condition just like Zasetsky's except restricted to colour words and ideas and it results precisely from damage to one restricted part of the temporal lobe. Damage in a different location leaves colour words intact but affects proper names. Damage in a third area leaves nominals (nouns and adjectives) intact but affects verbs, in a fourth and the capacity to structure sentences is affected (Damasio & Damasio 1992). The mediation between word-classifiers and thing-classifiers seems to be laid out with all the functional specificity of a computer's motherboard. Though it is easy to accept that there are links between word classifiers and thing classifiers it is not so obvious how to define the linking relation, yet it turns out we can make a definition in terms of classifiers. A Fodorian cow-tokeniser is a mechanism that says "cow s' when it sees cows, and flashes a light on top of its head when it hears "cow". If an utterance of " c o w " is made near to it, then its "cow"-dassifier detects it and flashes the light. If a cow is nearby, its cow-classifier detects it and delivers an utterance of "cow". So we have two classifiers. They are linked, for the verdict of the cow-classifier (the utterance) will in turn be detected by the "cow"-classifier with the result that the light flashes. Observe that this means that the "cow'-classifier is also a cow-classifier, and with the same visual presentation property. Let's set this out
TOWARDS A WINOGRAD/FLORES SEMANTICS
75
more carefully using the (P, o-, v) formulation of the last section. The tokeniser comprises two classifiers: M which is a (cow, visual, utter)-classifier and N which is a ("cow", audible, flash)-classifier. But it turns out that N is also a (cow, visual, flash)-classifier, and for this reason we s a y that M and N are linked. More abstractly suppose M is a (PM, °M, vM)-classifier and N is a (PN, O'N, VN)-Classifier within some class of situations S. Then M and N are linked provided that we have either M is (PN, ON, vM)'classifier or N is a (PM, crM, vN)'classifier. Notice that linking thus defined is reflexive and hence that linking is not a causal relation. Ruth Millikan's (1984) hubots have a system of classifiers and links integrated into social behaviour. When a hubot comes across gold, it calls out "golper" and other hubots that hear this "golper" come running to get their share. These other hubots hear a call, classify it as "golper" and initiate a search that terminates when they find a hubot - it does not have to be the original speaker - and classify the stuff it is holding as golper. So in these situations we can take the "yes"verdict of the golper-classifier as indicating the previous utterance of "golper". The golper-elassifier is a "golper"-classifier as well, and so they are linked. With Asimov in the workshop things are much the same. When George utters " G e t me a hammer" in Asimov's hearing, Asimov will classify the word " h a m m e r " and then go off looking for one, a search that will terminate (usually) when he classifies an object as a hammer. In these circumstances the hammerclassifier is a "hammer"-classifier as well, and with the same presentation property. So they are linked. Finally consider George himself playing a game of naming things in the workshop. He points at a tool and says " h a m m e r " or "chisel" or whatever, depending on what he takes the tool to be. The recognising of hammers in front of him (presentation property) is done by his hammer-classifier M. He recognises his own words when he utters " h a m m e r " , so his "hammer"-classifier N will then give its "yes"-verdict too. In these circumstances N is also a hammer-classifier with the same presentation property as M. So we say that M and N are linked. Of course this naming game is hardly something that people do too much, but that is not the point. If a person could not do it, could he still be l i n k i n g - o r " m e a n i n g " - h a m m e r s by " h a m m e r " ? Was not the unfortunate Zaxetsky unable to play exactly the naming game? Of course not all word-classifiers are linked to thing-classifiers- words like " o f " , " b u t " or " a n d " are not. The words that are we shall call, using the medieval term (Broadie 1987), categoramatic words.
4. Co-Ordination by a Means-the-Same-by Relation We now want, using idea of a system of classifiers and links, to construct for a categoramatic word ~ a relation X and Y mean the same by ~ in situation s. With the apparatus of classifiers in links in place this can be done quite quickly. Taking " h a m m e r " as an example firstly we require that within the situation s, both X and
76
P E T E R MOTT
Y have "hammer"-classifiers which they link to some thing-classifier. For X this means:
(1)
(3P)(3M, o-M, ~,M)(3N, O'N, UN)(M is a ("hammer", O'M,vM)-classifier in X & N is a (P, o N, VN)-Classifier in X & M links N)
The key point here is that to understand "hammer" requires one to link it to some classifier for some property P, not necessarily to the property of being a hammer. The formulation is rather complicated, but since the classifiers, methods of presentation, and verdicts in (1) are all existentially quantified we may simplify it by putting: "¢ links P in X " for "(3M, ~rM, VM)(3N , ON, VN)(M is a (¢, ~rM, VM)classifier in X & N is a (P, ON, vu)-classifier in X & M links N)". Then (1) so abbreviated is: (2)
(3P) "hammer" links P in X
Placing a system of classifiers and links for "hammer" in both X and Y then gives us: (3)
(3P)("hammer" links P in X) & (3Q)("hammer" links Q in Y)
Both now link their "hammer" classifiers with a property classifier. Following our original outline, for meaning the same by "hammer" X and Y must classify the same things as hammers, and inspection of (3) shows us that this requires that the linked classifiers must classify the same property, at least throughout the situations S where they mean the same by "hammer". Thus we have: (4)
(3P)("hammer" links P in X & "hammer" links P in Y)
If (4) is true in situation s then in that situation X and Y mean there the same by "hammer". What it is that they mean by "hammer" we do not try to say (that way lies the bog of reference), though assuredly there is something as our existential quantification shows. In different situations, of course, there may no longer be any common property or type linked to "hammer" in X and Y and then X and Y will no longer mean the same by "hammer". If Asimov classifies things as hammers by their green handles and George classifies them as we do, then just so long as all hammers in their workshop have green handles they will classify the same things as hammers and their co-operation can go ahead. They can mean the same by "hammer" because they will in the workshop link "hammer" to Ax(green handle(x) & hammer(x)). Such a consensus would be fragile. To make it more solid it would be necessary to design Asimov to co-classify with George over a wider range of examples of hammers. But on the present theory there is no qualitatively different approach available. Two friends, for example, may believe for years that they share ideas as to what democracy was only to find their agreement vanish during some major social upheaval. Being humans they would then no doubt try to re-negotiate the meaning of "democracy" between themselves to secure consensus - maybe they would succeed, maybe not. The point is
TOWARDS A WINOGRAD/FLORES SEMANTICS
77
that agreement over meanings is always provisional, more or less fragile, and has actively to be maintained. The means-the-same-by relation of (4) has the expected formal properties. Firstly, if a categoramatic word ~ means something for X (there is a system of classifiers and links in place) we can conclude that X means by ~ what X means ~. So the relation is reflexive. By inspection of (4) it is obviously symmetric; if X means by ~ what Y means by ~ then conversely, Y means by ~ what X means by ~. Transitivity is subject to a condition of no a m b i g u i t y - that a categoramatic word links to just one property. Otherwise X could use a word with one link (i.e. in one sense) when talking to Y and with a second link when talking to Z (he may even be unaware that he is doing this) so that though he could talk to both they could not talk to each other. Assume unique links and transitivity follows.
5. The Disjunction Problem According to what Fodor (1987) calls the Crude Causal T h e o r y " c o w " means cows because cows cause us to token " c o w " to ourselves. This causal correlation is not perfect. When an animal is a long way away or the light is bad we sometimes token " c o w " in the presence of a bullock. H o w are these errors to be understood? The problem is that the supposed misclassification of a bullock as a cow need not be taken as such at all. Instead it can be used as evidence that " c o w " means cow-or-bullock, for the causal correlation between tokenings of " c o w " and cow-or-bullock is better than that between " c o w " and cows. E r r o r then becomes impossible, because meanings expand to accommodate it. This is the Disjunction P r o b l e m ) The causal-informational theorist in seeking a solution to the Disjunction Problem must preserve his basic claim that "cow" really means cows because of some causal connection between cows and " c o w " tokenings. We do not make any such claim. In our theory there is no question when "cow" really means cows, there is only the question when two or more speakers mean the same by "cow". F r o m this perspective sentences like (A)
" C o w " means cows
turn out to be analytic. They assert no empirical relation between words and things at all, but instead tell us just that " c o w " is a categoramatic word. Consequently a sentence like: (B)
" C o w " means cow-or-bullocks
will be analytically false just because it is not compatible with (A). The contrast that the causal-informational theorist seeks to draw is therefore correct. T h e r e really is a fact of the matter that makes (A) true and (B) false, but it is not a fact about a causal connection between words and cows. It is an internal fact about the words of the language. Our theory is not developed enough to demonstrate this claim rigorously, but
78
P E T E R MOTT
we sketch an argument all the same. Let X be a speaker of English and let us enquire under what conditions X will believe: (1)
"gorse" means furze.
This certainly requires that X understand the categormatic word "furze" that is used in the sentence. X must also understand "gorse" as another categoramatic word, for though the word is only mentioned not used in (1), anyone who believes (1) must also believe .... Furze" means gorse", and that does require understanding the categoramatic word "gorse". So X must, according to our theory, have classifiers and links such that: (2)
(3P)("gorse" links P in X) & (3 Q)("furze" links Q in X).
On its own (2) says nothing about synonymy. For that we need to make a further assumption that synonymy for categoramatic words arises when both words are linked to classifiers for the very same property. Thus from (2) we obtain the following to represent X's belief that "gorse" means furze: (3)
(3P)("gorse" links P in X & "furze" links P in X).
Note that this does not distinguish the belief that "gorse" means furze from the belief that "furze" means gorse which is as it should be. The word "cow" is categoramatic in English so that for X to understand it we require that it be linked to a classifier: (4)
(3P)("cow" links P in X).
This is logically equivalent to: (5)
(3P)("cow" links P in X & "cow" links P in X).
But by (3) this is the condition for X believing that: (6)
"cow" means cows.
The conclusion that X believes that "cow" means cows is thus a trivial consequence of the fact that "cow" is for X a categoramatic word and holds for all such words. "Word" means words, "life" means life, "ink" means ink and so on. All these are true merely because the involve categoramatic words. Correspondingly such formulations as "cow" means cow-or-bullock and "ink" means ink-or-paper are all false because incompatible with the foregoing. The above is a solution to the Disjunction Problem (it is not a complete solution because I have not explained the notion of incompatibility which it trades on). It shows, I hope, that the approach adopted here provides a genuine alternative to the causal/informational view. The question of error remains outstanding, for our theory is not yet developed far enough to give an account of that. We will, however, suggest how error might be accommodated in a fuller version of a Winograd/Flores semantics. Consider a group of people who all co-classify objects as being hand-written
TOWARDS A WINOGRAD/FLORES
SEMANTICS
79
letter 'A's until a new style of calligraphy is proposed by one of them. It turns out that the community divides, half classify the proposed new 'A's as such while the other half do not. Both groups are perfectly sincere. Those who classify it as an 'A' see it as such, those who don't see it as something else. On our account the underlying situation is quite clear. The 'A'-classifiers in that community coincided until the range of situations was widened by the calligrapher's invention. In the wider class this consensus is lost and the two groups no longer mean the same by " a n A " . Is one half of the community wrong and the other half right? The present theory provides no resources for a decision. A Peircean (I suppose) will enquire which group eventually triumphed and say that one was right. A Realist will invoke the Law of Bivalence to the effect that one group is right and the other wrong even if there is no telling which (unlike the Peircean the Realist will concede that the entire community could reach the wrong conclusion and be forever wrong on what an 'A' was). Peircean and Realist intuitions have to be acknowledged, but for us there is no question of error here at all. We simply have the breakdown of consensus due to changing situations. Neither group is in error, they just mean different things by "an A " . Error requires a more substantial context. So what might be added to the picture to supply such? Suppose now that the new 'A's appear not on the end of a calligrapher's pen but instead in an old document. Assume that only a single individual at first accepts these marks as 'A's, but she is a scholar and she uses her assumption together with extensive other knowledge to translate the document, and her translation is plausible and wins a general acceptance within the community. They will therefore have to agree that the marks are 'A's, even if they don't look like 'A's. Agreeing that, they will acknowledge previous error. The situation arises because they have available two ways of classifying these particular marks, a way based upon the way the marks look and another way based upon the conclusion that the translation is correct. The two ways conflict and so a choice has to be made between them. The presence of that choice is what makes error possible in this case and thus what distinguishes the second story from the first. Classifiers and links are not enough on their own to allow the possibility of error. For that you need systems of classifiers and links inter-connected so that conflicting classifications become possible. Error can only appear when such a system is in place. Our present theory does not go that far, so it does not explain error, though it does indicate in what direction an explanation should be sought. It may also throw some light on why the Disjunction Problem has proved so recalcitrant for the causal/informational theorist. He may just be using a framework too impoverished to support the phenomenon he wishes to explain.
6. Winograd/Flores and the Objective World Patterns of co-ordinated behaviour Winograd/Flores call, following Maturana, "consensual domains". Clearly numerous patterns of human behaviour comprise such domains. Take a dance as an example:
80
PETER MOTT
Perhaps the most famous and stately dance of this period was the Pavane (of Spanish origin) which is very fully described in Tabouret's Orchdsographie, the earliest work in which a dance is found minutely described. The Pavane which was really more of a procession than a dance, must have been a very gorgeous and noble sight, and it was perfectly suited to the dress of the period, the stiff brocades of the ladies and the swords and heavily plumed hats of the gentlemen being displayed in this simple and dignified measures to great advantage. (Encyclopaedia Brittanica, p. 16, vol. 7, 1926 edition) In the Pavane there are rules to be followed, things to be done at certain times, m o v e s of considerable intricacy to be made at the right m o m e n t , positions to be a d o p t e d and then changed in response to the other's move and so on and, of course, you don't Pavane alone. W h a t e v e r the rules might be, they are not defined for one player. Dancing is a social, co-ordinated activity, a consensual domain. Social life is comprised by a large n u m b e r of such consensual domains. W i n o g r a d and Flores take language itself to be one. Language is a consensual domain: but it is m o r e than that for it is used to organise and co-ordinate other consensual domains. It is this power of language to organise and co-ordinate other social activities that sets it apart and gives it meaning. This is, as we said at the start, the root of the W i n o g r a d / F l o r e s conception of meaning. Consider a bird alarm call. A flock is feeding on the ground, one sees a p r e d a t o r , gives an alarm call and they all fly away. The action of all flying away is co-ordinated; the alarm call is what achieves this co-ordination and hence has a meaning. It means, m o r e or less, " S c r a m ! " (compare Millikan (1989) on the behaviour of beavers). If the birds systematically no longer heeded the alarm call then it would simply cease to be an alarm call and to have meaning amongst t h e m - even if its causes remained precisely the same. Again, consider G e o r g e and Asimov. W h e n G e o r g e wants a h a m m e r he calls " H a m m e r ! " , and A s i m o v goes and gets one and their work goes ahead. It is no good if Asimov responds to " H a m m e r ! " by nodding, or shouting "Chisel!", or going off and repacking himself into his delivery crate. In a very obvious sense that is a breakdown. Winograd (1987) offers a m o r e subtle example: A friend walks into the kitchen where person A and person B are sitting and asks "is there any water in the refrigerator." A says "Yes" and B says "No". The friend looks in and says "I don't see any." A responds, "In the cells of all the vegetables." T h o u g h A here tells the literal truth his response causes a b r e a k d o w n in co-ordination, for the friend sets out to get a glass of water and fails. Breakdowns are occasions when language fails to co-ordinate activity. If they b e c o m e frequent enough, people stop talking. N o w the question here is why such b r e a k d o w n s are not ever so frequent. H o w is it that when G e o r g e asks Asimov for a h a m m e r he regularly gets a h a m m e r and not something else? H o w is it that when birds call out other birds recognise alarm calls, or when friends ask whether there is any water in the refrigerator we are able to give t h e m the right answer? It cannot be a coincidence, something must secure this reliability. Our answer has been that the reliability is explained
TOWARDS A WINOGRAD/FLORES SEMANTICS
81
because we have classifiers that co-classify. They co-classify because they detect within a given range of situations the same types or properties of objects. But in saying this are we not tacitly returning to the "Rationalist Tradition" which Winograd/Flores repudiate? I do not think so, for the following reason. It is the not the existence of types that Winograd/Flores dispute but rather the existence of an antecedently given fixed set of types and we have no need to assume any such set. They deny that " A language can be arbitrary in using the words "dog" and "on" or "chien" and "sur", but it is constrained by the nature of the world to group a certain set of objects or properties together under whatever names it uses" (61). We can echo their denial for our theory gives no assigned limit to what classifiers may classify. Winograd/Flores early in their book recall the well-known fact that when a stick is illuminated by white light from one side and red from another it casts two shadows, one pink and the other green, though there is no light of the wavelength of green light present. So when we classify something as green we are not in general classifying wavelengths. Our theory is comfortable with this. We can have a classifier that behaves in this way (the mechanisms involved seem to be quite well understood), and it can be linked to a "green" classifier. Installed in Asimov such a system should enable him to mean what we mean by "green". More than that our theory cannot supply, but more than that as theorists of communication we do not need. We cannot be accused of their fallacy of a fixed objective world. However they appear to go further: " . . . language does not describe a pre-existing world, but creates the world about which it speaks." They give an example. "There are whole domains such as those in financial markets involving 'shares', 'options', and "futures' whose existence is purely linguistic-based on expressions of commitment from one individual to another" (174). Even more abruptly "From these points we are led to a more radical recognition about language and existence: Nothing exists' except through language" (68, italics in original). Such claims unsympathetically construed will seem excessive, but they can be taken as additional articulation of the denial that there is a single fixed world. The story might go like this. There is a sense in which an elephant and an ant passing across the same bit of forest still inhabit completely different worlds. Because elephants are a lot bigger than ants they meet with, and discriminate, a wholly different set of objects. It is not only size that makes a difference. A baboon that can pick up and manipulate objects will inhabit a different world from a wildebeest that can only tread on them, even though they are within a hundred yards of each other. Different sensory capacities also define different worlds and so also, at least for humans, do different languages and theories. Worlds can diverge in varying degrees. The worlds of the elephant and the ant diverge radically, those of the baboon and the wildebeest rather less so, people who differ only in language and culture less so again. But there is still sense in saying that they inhabit different worlds and there is sense in saying that language contributes to delimiting the world they ,inhabit.
82
PETER MOTT
On the other hand in setting out this example we apparently assumed there was only one world, from which all these particular worlds were carved out. To try and relieve this tension, let us start with a universe, or "objective background" as I shall call it, of all the things and types of things there are (not possible objects). The objective background supplies not only solid spatially bounded medium size objects like billiard balls, but also explosions, football matches, clouds, the sky, whirlpools, shadows and so on. Because some organisms make things it includes molehills, wasp nests, the Taj Mahal, and the market in financial futures. Exploiting an analogy with Set Theory where no set includes every set, indeed no set even manages to include all the ordinals, we say that no world includes all the things in the objective background. So the objective background is not itself a world, in particular it is not our world. But just as in Set Theory (at least realistically interpreted) all the ordinals exist, so do all things exist in the objective background. Organisms construct worlds out of the objective background as a result of the way they interact with that background. A n organism's world is the set of objects that it discriminates and interacts with. The elephant and the ant do not inhabit the same world, but both their worlds are carved from the objective background. There are such things as financial markets and cocoa futures. They are part of our world and also part of the objective background. Reverting to their example, Winograd/Flores said that what exists are 'futures' and 'shares', but these are just words which existed long before there were either shares or futures. There are shares as well as 'shares' and futures as well as 'futures' and the Taj Mahal as well as the 'Taj Mahal' and to deny this is to adopt a sort of linguistic Idealism which in fact they explicitly dissociate themselves from (68). To maintain the distinction between 'shares' and shares while keeping the intuition that there is not unique ready-made world, some such device as the objective background seems essential. That is all we require for our classifiers. The above really finishes our defence against the charge of departing from the spirit of Winograd/Flores, but there is one more issue that I would like to explore briefly. Take the phrase "top athlete". You and I may link that phrase to classifiers that co-classify in our situation so that we call the same people top athletes. It is certainly possible that we spend more time talking in general about top athletes than we do actually classifying particular individuals are such. We may discuss at length, for example, whether it is harder for a woman than a man to achieve recognition as a top athlete and in this discussion we just assume the class of top athletes as agreed between us (though to be sure examples must occasionally come up). Surely this talk could survive even if our situation changed such that we no longer classified the same people as top athletes even though we continued to believe that we did. In which case an anchor point to the nonlinguistic world would have silently come adrift without our noticing and without our discourse being apparently affected at all. One wonders whether there could not be bodies of discourse where this has happened more completely, discourse
TOWARDS A W I N O G R A D / F L O R E S SEMANTICS
83
where there is no longer any clear connection to the non-linguistic world remaining, where speakers expected certain responses and moves to be m a d e , as in a game. This discourse would no longer serve to co-ordinate any action at all except in one very special case. It would perforce contribute to the maintenance of itself. Such a discourse would still be a consensual domain, but in exactly the sense that the dignified Pavane was. It could still be socially important (as the rituals of the Pavane might well have been). For this sort of linguistic activity, however, there would be no semantics in the present sense.
Appendix: A Situation-Theoretic Model of the Theory of Classifiers. T h e object of this appendix is to develop some Situation T h e o r y so far as is n e e d e d to present m o r e precisely our theory of classifiers. The treatment aims to be m o r e or less serf-contained but is very far from comprehensive, Devlin (1991) is a good general introduction. T h e basic idea of Situation T h e o r y is of a situation s supporting (~) an item of information or infon ~r. Infons are states of affairs in the Logical Atomist tradition. The first mention of one that I know of occurs in the Introduction to Russell (1910) where he stays: The universe consists of objects having various qualities and standing in various relations. Some of the objects that occur in the universe are complex. When an object is complex, it consists of interrelated parts. Let us consider a complex object composed of two parts a and b standing to each other in the relation R. The complex object "a-in-the-relation-R-to-b" may be capable of being perceived... (Chapter II). T h e infon that Russell describes would be represented in the current notation as ((R, a, b, 1)}. If in situation s the object a does stand in the relation R to the object b then s~((R,a,b, 1)); if it !!does not so stand then s~{R,a,b,O)). Russell had only one s i t u a t i o n - the U n i v e r s e - which settles everything that can be settled. Instead of the Universe, m o d e r n Situation Theory has m a n y partial situations. In fact a situation s m a y support just a single infon ~r or even none at all. It m a y be that for any object x at all, if s ~ ((P, x, 1}) then also s ~ ((Q, x, 1}). In that case s supports the constraint that being P involves being Q, which we write P ~ Q. Classifiers obviously involve constraints, for what makes something a P-classifier is the constraint between presenting a P-object and delivering a "yes"-verdict. Setting up some Situation Theory to represent this rather complicated constraint is the first thing to do, then we prove some simple propositions about classifiers. We assume a fixed set P A R of parameters, a set P O L = {1, 0} and sets I N D , REL: I N D is a denumerable set of objects or individuals. R E L is a set of relations of various arities including a distinguished binary relation ~ .
84
P E T E R MOTT
All these sets are disjoint. Bold face will be used for the e l e m e n t s of P A R which m a y be t h o u g h t of as place-holders (see Devlin 1991). By ( ( a x , . . . , an)) w e m e a n the tuple ( a l , . . . , an)). By a ^ b we m e a n (2, a, b ) and by [ a l b ] we m e a n (3, a , b ) . N e x t we define by simultaneous recursion two sets I N F O N and T Y P E o v e r I N D and R E L as the smallest sets such that: a) if R is an n-ary relation in R E L , a l , . . . , an are in I N D U P A R U R E L , and p @ P O L t h e n o- = ((R, a l , . . . , an, p )) is in I N F O N . T h e free p a r a m e t e r s fr(o-) of 0- are fr(al) U - - - U f r ( a . ) , w h e r e fr(a) = {a} if a E P A R and otherwise is e m p t y . b) if 0-1 and 0-2 are in I N F O N then so is 0-1 ^ 02. T h e free p a r a m e t e r s are given b y fr(0-1/x 0-2) = fr(0-1) U fr(o-2). c) if 0- E I N F O N and a l , . . . , a n are in P A R then [ a l , . . . , a n ]0"] ~ T Y P E . T h e free p a r a m e t e r s are fr(0-) - { a l , . . . , an}. If n = 1 then T is a unary type. d) ff T = [a 1. . . . , an 10"] @ T Y P E and a l , . . . , a n are in I N D U P A R U T Y P E , and p E P O L , then ((T, a l , . . . , an, p)) E I N F O N . T h e free p a r a m e t e r s are fr(T) U fr(a,) U - - . U fr(an). G i v e n an infon 0-, the dual ~, of 0" is the infon o b t a i n e d f r o m 0- by exchanging polarities. A n infon is parametric if it includes a free p a r a m e t e r , otherwise it is non-parametric. T h e e x a m p l e s b e l o w illustrate these definitions.
Infon
Meaning
({likes, john, red, 1))
John likes the colour red
{{red, a, 0))
i s not red
{([a] red, a, 1))], m, 0))
mis not red
The relation red can occupy an argument place in an infon Dual to ({red, a, 1)). a is an individual parameter. The infon is parametric. [a I {{red, a, 1))] is the type of red things. Infon is non-parametric.
A situation s o v e r I N D and R E L is a set of n o n - p a r a m e t r i c infons o v e r I N D and R E L such that: a) N o infon and its dual are b o t h in s. b) o- A ~- E S iff 0- E S and ~- E s. c) ( ( [ a l , . . . , a n l 0 - ] , a x , . . . ,an, 1)) E s iff 0-a~. . . . . . E s w h e r e 0-a~. . . . . . n c o m e s f r o m or b y replacing each occurrence of a free p a r a m e t e r a i in o- by a r ( T h e p a r a m e t e r s a i n e e d not occur 0-.) d) ( ( [ a l , . . . ,anl0-], a l , . . . ,an, O))Es iff Kal . . . . . . n e S " e) L e t P, Q be u n a r y types or relations without free p a r a m e t e r s . T h e n ( ( ~ , P , Q, 1)) E s i f f f o r every x ~ I N D , if ((P,x, 1)) E s then ((Q,x, 1)) ~ s . ( W e write (( i f , P, Q, 1)) as P ~ Q and call these infons constraints). A situation s supports an infon 0- (s ~ 0-) iff 0- E s. A s an e x a m p l e of these definitions, note that by (c) we m a y express that an o b j e c t m is red in s in two equivalent ways: s ~ ((rd, m, 1)) or s ~ (([a[ ((red, a, 1)}], m, 1)). But then by (e), s must s u p p o r t a constraint: s ~ r e d ~ [ a ] ((red, a, 1))].
TOWARDS A WINOGRAD/FLORES SEMANTICS
85
Let P be a property or unary type without free parameters. We define the type of being a P-classifier o f o-objects by verdicts v (a (P, ~r, v)-classifier). First construct the two constraints: zl = [a[ <
> ^ crI ~ [aIv] z2 = [a I ((n, a, 0)> A o-l ~ [a I gl Now letting (P, o-, v) be zl ^ z2 the type of a (P, o-, v)-classifier is:
[m I(P, o-, v)] There is no requirement that any parameters actually occur free in or and v. The expectation is that m and a will occur in o-, and that m will occur in v. If not the property of being a (P, ~r, v)-classifier just won't be very interesting (see Proposition 4), but it will still be correct. Also to simplify the following we have assumed that the "yes"-verdict and the "no"-verdict are dual infons. Finally we define: m is (P, or, ~,)-classifier in s iff s ~ (( [m I (P, o-, v)], m, 1 )). This is the situation theoretic version of the notion of classifier developed in the main part of the paper. We finish with some simple propositions about how classifiers behave. Firstly we shall show that when a P-classifier is presented with a P-object it says "yes" and when presented with non-P object its says "no". The infon cr represents the presentation property of a (P, g, p)-classifier, and ~rm,. is the result of replacing the parameters a and m by the individuals a and m. P R O P O S I T I O N 1. I f m is a (P, o, v)-classifier in s and s ~ o%,~ then (i) if s~ ((P,a, 1)} then s ~ v . . . . (ii) if s ~ ((P,a,O}) then S~ ~m... We prove (i). Because s e (([m] (P, (r, v)], m, 1)) we know: s ~[a I ((P, a, 1>> A Crm]~ [a I v.,]
s [.I <
> ^ m] [al We assume that s ~ crm,. and s ~ (( P, a, 1)) so that s ~ (( P, a, 1)} A O'm,.. Hence s ~ (([a t ((P, a, 1} A or l, a, 1} and so s ~ (([a[ urn], a, 1)) and s ~ vm, a. For (iX) a dual argument suffices. [] There is no direct converse to Proposition 1. We cannot show that when a P-classifier is presented with an object a and says "yes" then a must be a P-object for there is the possibility that the situation simply does not settle whether a is a P-object at all. We do, however, have the following partial converse: P R O P O S I T I O N 2. I f m is a (P, o-, v)-classifier in s and s ~ ~rm,a then (i) if s ~ Vm,a then not s ~ ((P, a, 0)) (ii) if s ~ g..,a then not s ~ ((P, a, 1))
PETER MOTT
86
For (i) we have s ~ (([m] (P, ~r, v)], m, 1}} and so:
(1)
s ~[a] ((P, a, 0} ^ O-m]~ [al gm] .
We assume that s ~O'm,a. Suppose s ~ ((P, a, 0}}. Then from the constraint (1), s ~ ~m,,- But this is impossible for s is a situation and cannot contain both an infon and its dual and s ~ v,~,.. Once more (ii) is just the dual of this. [] Of course if situations were taken to settle all facts then this complexity would not arise as we could infer s P 6 from the failure of s ~ o-. Next we shall show that classifiers are extensional in that if two properties P and Q are exemplified by the same objects in a situation then m will classify both or neither. Assume that for every x ~ IND we have: s ~ ((P, x, p)) iff s ~ ((Q, x, p)). Then: P R O P O S I T I O N 3. m is (P, or, v)-classifier in s iff m is ( Q , ~r, v)-classifier in s. Suppose m is (P, o', v)-dassifier in s. Thus:
s~[al ((P, a, 1)) A O-m]~
[alvm]
s~[a] ((P, a, 07) A O-m]~ [a ] g,~] Now assume s ~ ((Q, a, 1)) ^ O'm,.. SO S ~ ((P, a, 1)) A O'm,.. Thus s ~ Vm,a. Since a was arbitrary we have:
s e[al (
vm].
The dual argument gives:
s~[al (
[]
Merely being a classifier (P, o-, v) is not interesting, what matters is the presentation property and the verdicts, and how they are integrated into a system of classifiers and links. In fact anything at all is a classifier of any property at all if you gerrymander the verdicts enough: P R O P O S I T I O N 4. There is a verdict such that rn is a (P, o-, v)-classifier. Let v = ((P, a, 1)). Then s ~ [ a I { P , a , 1)) A o-,n]~[a] urn] trivially and likewise s ~ [a[ ((P, a, 0)) A cry] ~ [a[ ~ ] . Thus m is a (P, or, v)-classifier. []
Acknowledgements I should like to thank two anonymous referees of this journal for criticisms of an earlier draft. These led to extensive reworking of the last two sections.
TOWARDS A WINOGRAD/FLORES
SEMANTICS
87
Notes i Parenthesised page references are always to Winograd and Flores (1986). 2 As a matter of fact Gorman and Sejnowski do not mention mines at all. They analysed sonar returns from "a metal cylinder and a cylindrically shaped rock positioned lengthwise on a sandy ocean floor". 3 Sterelny (1990) provides a good account of attempts to solve the Disjunction Problem in the causal/informational framework. Fodor (1984) introduced the problem. His preferred solution he calls the Asymmetric Dependence Theory. This is introduced in his (1987) and further discussed in his (1990).
References Bechtel W. and A. Abrahamsen (1991), Connectionism and the Mind, Oxford: Basil Blackwell. Broadie, A. (1987), Introduction to Medieval Logic, Oxford: Clarendon Press. Churchland, P. M. (1988), A Neuro-Computational Perspective. The Nature of Mind and the Structure of Science. Cambridge, MA: MIT Press. Damasio, A. and H. Damasio (1992), 'Brain and Language', Scientific American 267(3), 63-71. Devlin, K. (1991), Logic and Information. Cambridge: Cambridge University Press. Fodor, J. (1983), The Modularity of Mind, Cambridge MA: MIT Press. Fodor, J. (1984), 'Semantics, Wisconsin Style', Synthese 49, 231-250. Fodor, J. (1987), Psychosemantics, London: MIT Press. Fodor, J. (1990), A Theory of Content & Other Essays, London: MIT Press. Gorman, R. E and T. 8ejnowski (1988), 'Learned Classification of Sonar Targets Using a Massively Parallel Network', IEEE Transactions, Acoustics, Speech and Signals Processing 36(7), 1135-1140. Luria, A. R. (1972), The Man with a Shattered World. Penguin. Millikan, R. (1984), Language, Thought, and Other Biological Categories. Cambridge, MA: MIT Press. Millikan, R. (1989), 'Biosemanties', Journal of Philosophy 86, 281-297. Putnam, H. (1981), Reason, Truth and History, Cambridge: Cambridge University Press. Putnam, H. (1983), Realism and Reason, Cambridge: Cambridge University Press. Putnam, H. (1987), The Many Faces of Realism, Illinois: Open Court. Russell, B. and A. Whitehead (1910), Principa Mathematica, vol. 1, Cambridge: Cambridge University Press. Sterelny, K. (1990), The Representational Theory of Mind, Oxford: Basil Blackwell. Winograd T. (1985), 'Moving the Semantic Fulcrum', Linguistics and Philosophy 8(1), 91-104. Winograd, T. and F. Flores (1986), Understanding Computers and Cognition, New York: AddisonWesley. Winograd, T. (1987), Three Responses to Situation Theory. Report CSLI-87-106, Ventura Hall, Stanford. Wittgenstein, L. (1953), Philosophical Investigations. Oxford: Basil Blackwell.