In this section I attempt to present a brief but coherent theoretical and methodological statement, but in doing so I find it necessary to race headlong through some hotly contested territory, imagining attacks from every theoretical orientation as I go. There is hardly a reader who will not find objectionable some position that seems to be taken without further justification in a casual phrase. I can only plead that I have tried to choose my phrases carefully and that the positions I adopt, though controversial, have been adopted and argued at length by others and could be defended.^1.2 Since one's theoretical stance tends to permeate one's work in subtle ways, I feel it is necessary to lay out the presuppositions of this enterprise at the beginning, even though an adequate treatment of these issues would require a volume in itself.

A scientific theory is a more or less formal explanation of more or less extensive data. Formality promotes intellectual honesty in what the theory predicts. Confirmation and falsification result from the comparison of the theoretical prediction with the data and tell us what data is and is not covered.^1.3 A theory is ultimately judged by the elegance of its formal explanations and the coverage of its data. Idealization is a way of trading in extent of coverage for elegance in explanation.

This paragraph merits repetition.

A scientific theory is a more or less formal explanation of more or less extensive data.^1.4 Formality promotes intellectual honesty in what the theory predicts. Confirmation and falsification result from the comparison of the theoretical prediction with the `data' and tell us what `data' is and is not covered. A theory is ultimately judged by the elegance of its formal explanations and the coverage of its `data'. Idealization is a way of trading in extent of coverage for elegance in explanation.

In presenting a theory of discourse, it is necessary to state at the outset what is to count as an explanation, what makes an explanation formal, and what data one is willing to be held responsible for. Each of these is taken up in turn.

 

Explanation in AI and cognitive psychology is based on the computer metaphor: whatever else it may be, the brain is at least a kind of computer. We seek to understand how psychological processes could be implemented in terms of the symbol manipulation operations of computability theory. Insofar as we succeed, we will say we have ``explained'' them. There are several reasons this account of explanation is compelling.

The first is simply that it is often useful to try to understand one complex system by comparing it with another. Analogies elucidate. This is particularly true when different aspects of the two systems are open to inspection. Large computer programs are among the most complex objects which (in principle) are entirely under our control, and they exhibit behavior that at least superficially resembles some aspects of intelligent human behavior. The analogy between cognition and computation is fruitful for this reason alone.

A second reason for adopting the computer metaphor is purely technological. We know quite well how the level of symbol manipulation can be implemented in electronics, and thus whatever success we have can at least lead to useful computer systems. As Terry Winograd has said, if in trying to build an airplane, we end up building a boat instead, we'll go for a sail.

But the primary motivation for the computer metaphor is the promise it holds out for ``reducibility''.

Science is organized by levels, a strategy that is successful probably because nature is organized by levels. There are several ways we can view these levels. First, we can view them as levels of description. Nature cannot usefully be described solely in terms of the motions of elementary particles. We have found it convenient to define or hypothesize larger-scale entities and to couch our theories in terms of them. We then try to account for the behavior of these entities in terms of the entities provided by the theory of the phenomena one or two levels down. Thus, chemists seek to understand in quantum theoretic terms why molecules react as they do.

We can also view the levels as levels of organization. That is, they are not merely convenient fictions that allow our poor, finite minds to understand what is going on. There is something in nature that actually corresponds to these large-scale entities and actually behaves approximately in the manner that our theories describe. The argument for assuming these things are really out there in the world is what has often been said: We should adopt the ontology implied by our most successful theories. The reality of the ontology is the best explanation for the success of the theory. Molecules, cells, tissues and organs, organisms, herds, and nations are not merely stories we tell. They really do exist.

Evolution has proceeded by levels of organization because these represent levels of competence. Each level is characterized by the achievement of stable forms, out of which larger structures can eventually be constructed. Molecules are stable forms constructed out of atoms; stars, rocks and cells are stable forms constructed out of molecules; multicellular animals, including people, are stable forms constructed out of cells; and social organizations are stable forms constructed out of people.

There is much controversy about the ways in which a scientific theory of one level is or ought to be ``reducible'' to that of another level. Most arguments against the reducibility of a higher-level science, or ``macro-science'', to a lower-level science, or ``micro-science'', take the following form: the macro-science and micro-science each require idealizations for their most elegant formulations, and there may simply be a mismatch between their idealizations. For example, physiology concerns itself primarily with the prototypical members of a species, whereas population biology must concern itself with deviations from the prototype (Dupré, 1983). It is not just that elegance in the macro-science would be sacrificed. Usually, computational complexity precludes the reduction (although the reduction of the thermodynamics of an ideal gas to statistical mechanics is a notable exception, where computational complexity does not preclude the reduction). Moreover, the structure and behavior of an entity at the higher level cannot be explained by describing only that entity at the lower level and not the environment with which the entity interacts.^1.5 Thus, one cannot describe the structure and life history of a rock on the basis of its mineralogy alone without reference to tectonic influences, and similarly, psychological phenomena depend on sociological phenomena as well as physiological principles. Finally, it is often theoretically impossible even to state important global properties of large systems in the terms provided by scientific theories of their components (Davidson, 1981; Moore, 1980). Such arguments show that laws of the macro-science cannot be replaced by complex statements in the language of the micro-science and proven as theorems from lower-level axioms. The macro-science generally concerns itself with emergent entities whose boundaries become very fuzzy when unpacked into the entities of the micro-science. Prediction does not become possible in the macro-science, resident on the laws of the micro-science. The entities of the higher level are generally very complex dynamic systems of entities at the lower level, and although gross regularities may be established, the fine details of higher entities and processes cannot be derived. We understand the underlying physics of rivers, hurricanes, and volcanoes, but we can't predict their behavior, except within very coarse limits.

But this would be a very strong form of reducibility, one that might be called ``replacibility''.

It is nevertheless obviously true that, say, geological processes are ``implemented'' or ``realized'' as complex chemical and physical processes, and how they are implemented is an important question in its own right. An answer to that question constitutes another kind of reducibility, a ``reducibility in principle''. This is stronger than Davidson's (1981) notion of ``supervenience'', which means only that there is no change in properties at the higher level without a corresponding change at the lower level. In this sense, intentional psychology is surely supervenient upon neurophysiology. But the situation in geology is quite different. We know, in principle, how the behavior of a geological entity, such as a river, depends on chemical properties, such as the viscosity of water and of various minerals, and on the laws of physics. We really do have a story to tell.

There is a stage that some sciences have gone through and others have not, that represents a qualitative advance. It happens when it is understood, at least (and generally no more than) in principle, how the entities and processes at one level emerge from entities and processes at lower levels. Geology passed through this stage in the 1960s with the widespread acceptance of plate tectonics. Before then, ever since the eighteenth century, explanation in geology bottomed out in a mysterious process of ``uplift''. Plate tectonics explained uplift in terms of underlying physical processes. At that point, geology became, in the sense of reduction I intend here, a ``reduced'' science.

In this sense, it is desirable for psychology to provide a reduction in principle of intelligent, or intentional, behavior to neurophysiology. Because of the extreme complexity of the human brain, more than the sketchiest account is not likely to be possible in the near future. Nevertheless, the central metaphor of cognitive science, ``The brain is a computer'', gives us hope. Prior to the computer metaphor, we had no idea of what could possibly be the bridge between beliefs and ion transport. Now we have an idea. In the long history of inquiry into the nature of mind, the computer metaphor gives us, for the first time, the promise of linking the entities and processes of intentional psychology to the underlying biological processes of neurons, and hence to physical processes. We could say that the computer metaphor is the first, best hope of materialism.

The jump between neurophysiology and intentional psychology is a huge one. We are more likely to succeed in linking the two if we can identify some intermediate levels. A view that is popular these days identifies two intermediate levels--the symbolic and the connectionist.

 

 

The intentional level is implemented in the symbolic level, which is implemented in the connectionist level, which is implemented in the neurophysiological level.^1.6 The aim of cognitive science is to show how entities and processes at each level emerge from the entities and processes of the level below.

The reasons for this strategy are clear. We can observe intelligent activity and we can observe the firing of neurons, but there is no obvious way of linking these two together. So we decompose the problem into three smaller problems. We can formulate theories at the symbolic level that can, at least in a small way so far, explain some aspects of intelligent behavior; here we work from intelligent activity down. We can formulate theories at the connectionist level in terms of elements that behave very much like what we know of the neuron's behavior; here we work from the neuron up. Finally, efforts are being made to implement the key elements of symbolic processing in connectionist architecture. If each of these three efforts were to succeed, we would have the whole picture.

In my view, this picture looks very promising indeed. Mainstream AI and cognitive science have taken it to be their task to show how intentional phenomena can be implemented by symbolic processes. The elements in a connectionist network are modeled closely on certain properties of neurons. The principal problems in linking the symbolic and connectionist levels are representing predicate-argument relations in connectionist networks, implementing variable-binding or universal instantiation in connectionist networks, and defining the right notion of ``defeasibility'' in logic to reflect the ``soft corners'' that make connectionist models so attractive. Progress is being made on all these problems.

Although we do not know how each of these levels is implemented in the level below, nor indeed whether it is, we know that it could be, and that at least is something.

At the present time our computational models of mental processes fall far short of what people are capable of. But this in no way constitutes an argument against the computer metaphor. Whatever the limits of the computer metaphor are, we have not begun to approach them yet.

Should the enterprise succeed, there is no reason to feel that it would diminish in any way our image of humankind, any more than the view of the human body as a chemical and mechanical device has devalued the body. A theory that did turn out to be dehumanizing would simply be wrong; it would contradict the facts as we know them. So far, however, the evidence has all gone in the other direction, showing us, if the brain is a computer, what a magnificent computer it must be. The computer metaphor does not pose a threat to humanistic psychology, for as Boden (1977) has pointed out, it is a variety of humanistic psychology. It allows us to view people scientifically not merely as exhibiting behavior but as taking action.

 

When we adopt the computer metaphor, we are adopting a particular, very stringent requirement for explanation in our theory. An explanation is a specification of computable procedures that will produce the behavior under investigation. If what we provide isn't computable, it isn't an explanation, only a prose description, and most of the point of appealing to the computer metaphor is lost. This requirement is what Wilensky has called ``procedural adequacy'' (Schank and Wilensky, 1977).

This requirement places those who study discourse on the horns of dilemma. On the one horn, we would like our theories to be faithful, empirically adequate accounts of the way people actually process discourse, the knowledge they actually use, and the goals they are actually driven by; on the other, we require computable procedures that actually solve the discourse problems we are faced with. In most cases, we can't have both. If we adhere to empirical adequacy and do not go beyond what our data warrants, we will not solve the problems, for too much is going on that is simply unobservable. We will be condemned to sterile theories. When we try to construct procedures that work, we are on shaky empirical grounds. Computability forces us to specify procedures to a level of detail beyond what is justified by solid evidence. It is for this reason that work in AI often seems ad hoc from a cognitive point of view.

There is at least a partial escape from ad hoc theories, that workers in AI have not always availed themselves of. This is to frame the explanations at as abstract a level as possible, while still retaining computability. Workers in AI sometimes^1.7 couch their theories in terms very close to the actual code of some implemented system. But, in spite of the power of the computer metaphor, the brain and the present-day computer are sufficiently different that a discussion at this level of detail is of no psychological interest, whatever its technological merits. In line with current practice in AI, the theory presented here is expressed in formal logic. It thus retains computability while being abstracted away from implementational details, such as data structures and serial versus parallel computer architectures. The most any AI, or other ``top down'' (cf. Dennett, 1978), approach in psychology can hope to do is to discover a possible explanation of behavior. By making the formalism maximally noncommittal on inessentials by using formal logic, we expand the range of possibilities the theory marks out. Rather than running counter to psychological reality, as often assumed, formal logic enhances the psychological relevance of one's theories.

 

Now the data--what is to be our concern and what is not. Psychological data is abundant. We are flooded with masses of it everyday as we interact with people, hear them talk, and observe their behavior. This psychological data is available to everyone, virtually without effort. Slightly less accessible are people's reports on this behavior--for language, such things as judgments about the grammaticality of a sentence or the appropriateness of a response, and clarifications, paraphrases and expansions of ambiguous or elliptical utterances. Finally, there are various sorts of more or less exotic, hard-to-obtain data, such as data on reaction times and eye movements, and results of fMRI studies. Ultimately a psychological theory will have to be responsible for all of this data, but today it is necessary to choose a portion of the data that seems significant and coherent and looks as if it will yield a reasonable theory. This inquiry will focus primarily on the middle of the three categories of data - specifically, on interpretation reports, people's reports on how they have interpreted a word, a phrase, or a larger stretch of text. We cannot use the most accessible class of data--utterances--for a theory of discourse interpretation, because it is not data about interpretation. It is about production. There are also problems with certain of the more exotic classes of data, as discussed below.

The use of interpretation reports is the genius--and the most significant contribution--of Chomsky's competence-performance distinction. It is not possible to build a science of utterances, because it is not possible to predict utterances. The mystery of choice intervenes. But it is possible to predict interpretation reports of a very limited sort, such as grammaticality judgments. Whereas a science of utterances would be a science of performance, a science of interpretation reports is a science of competence. It is a fundamental assumption of modern linguistics that the knowledge of language that we make use of in producing and interpreting utterances is the same knowledge of language that we make use of in interpretation reports. This assumption, that there is such a thing as linguistic competence, is what makes prediction, and hence a science of language, possible.

Our primary data will thus be certain kinds of interpretation reports. The first are reports of what a referential expression refers to, such as that ``he'' in the text

 

John can open Bill's safe. He knows the combination.

refers to John and that ``the index'' in

refers to the index of the book John picked up. Second are brief expansions or paraphrases of such constructions as compound nominals and similar predications conveyed in sentences. Examples are reports that ``wine glass'' means a glass whose function is to contain wine, and that ``turned to the index'' in (1.1) means turning the pages of the book until finding the index, rather than, say, turning one's body to face the index.^1.8 Of course such reports are themselves only texts, but we will take them to be reliable and priveleged. A satisfactory theory cannot claim that a listener resolved an occurrence of ``he'' to Bill when the listener reports that he resolved it to John.^1.9

We will also consider as psychological data to be explained the very fact that people can interpret a text and respond in an appropriate way. For example, suppose I ask someone if he could hand me a pencil, and he hands me the pencil, rather than, say, replying ``Yes'' or staring at his hand in confusion. This is psychological data, and we may ask how such a thing could occur.

Precisely what phenomena, and consequently what sort of interpretation reports, I will attempt to account for emerges as the book progresses. But it is perhaps useful for me to admit a bias at this point. Psycholinguists are frequently concerned with discovering and modelling people's shortcomings. Among linguists this concern takes a more theoretical turn; so that their postulated mechanisms will not have capabilities people lack, they strive to constrain the power of their theories. The typical result among linguists and psycholinguists has been to show that a particular process is too powerful and fails to explain what people don't or can't do. By contrast, the typical result in AI shows that a proposed process is not powerful enough and thus doesn't explain what people can do. It is a frequent charge among linguists that the hypothesis of a particular mechanism is vacuous because it has the power of a Turing machine. But it seems completely obvious to me that a mechanism capable of understanding natural language discourse must have the power of a Turing machine. The question is what that Turing machine is. In fact, it is not even clear that a Turing machine is adequate; the computer metaphor really might not turn out to be appropriate. It is precisely this hypothesis that AI and cognitive psychology have taken it upon themselves to test. It is certainly the ultimate aim of any theory to account both for people's abilities and their inabilities, but it seems to me that people's ability to engage in discourse so vastly outstrips our ability to model it that the primary problem is not to constrain our theories, but to enhance their power.^1.10

Interpretation reports come in two varieties--positive and negative. Some interpretations do not occur. For example, in (1.1) the listener will not normally take ``the index'' to refer to the index of the first book mentioned in the bibliography of the book John picked up. An adequate theory must explain both kinds of judgments. But when it comes to the interpretation of discourse, it might seem that negative judgments are hard to come by. Interpretation depends heavily on the context of utterance and the knowledge that is shared by the speaker and listener. This leaves us a lot of room to maneuver. If we exclude codes, it is unlikely that ``the index'' in (1.1) will be taken to refer to John's pet cat, but perhaps if we were clever enough, we really could load the context in a way that would support such outlandish interpretations.

Suppose we include among our interpretation reports judgments about the acceptability of question-answer pairs. What would we say about the following pair?

 

A: Was he an opera buff?
B: No, he was in the war.

The coherence of these utterances is likely to elude us. This was actually overheard, however, in the following context. A was whistling an aria from an Italian opera. B remarked that his father often used to sing that, and then the above pair. So the interpretation of B's reply is that B's father learned the aria while he was stationed in Italy during World War II. People, especially linguists, are very good at imagining contexts that will turn the most bizarre collection of utterances into coherent discourse.

This might seem to pose a dilemma. Either our theory predicts a single interpretation, in which case it will be wrong in many contexts, or it predicts a (rather large) family of interpretations, in which case it is very nearly vacuous.

In fact there is no such dilemma. The interpretation procedure - call it --is a function not of the text alone but also of some representation of the context. Texts are interpreted with respect to a knowledge base, and much of what we think of as contextual differences can be characterized as differences among knowledge bases. The knowledge we bring to bear on texts includes knowledge of the local environment and the situation of utterance, knowledge of the surrounding discourse, knowledge of what the speaker's and one's own aims are, knowledge of what knowledge one shares with the speaker, as well as general world knowledge. The order of access of this knowledge may vary according to salience and other attentional factors. Each of these aspects of knowledge will be examined in the chapters that follow. For now let us encapsulate it all in the symbol . Then we can summarize interpretation in the following ``formula'':

An interpretation procedure is applied to a text and a knowledge base to produce an interpretation .

This formula says that one cannot talk about the interpretation of a text without specifying the knowledge base that the text is interpreted with respect to. To put it another way, the context is one of the parameters of interpretation. If the context is changed, is changed, and we would expect to produce a different result. This is not a way of squirming out of the requirement of falsifiability. We have not thus made our interpretation process ``tailorable'', in the sense of van Lehn et al. (1983), to virtually any answer we desire. Once the new is specified precisely, the requirements on the theory are still as stringent: the anomalous interpretation must be produced. Texts really are interpreted differently in different contexts, and this is a fact that requires explanation in discourse theory.

The situation is quite analogous to one Lakatos (1970) imagines in prerelativistic celestial mechanics. An astonomer uses Newtonian mechanics and some observed initial conditions of a planet to predict its orbit. If the prediction fails to be confirmed, it does not lead him to reject Newtonian mechanics. He is more likely to postulate an unknown planet near the known one which perturbs its orbit. Similarly, if our interpretation procedure fails to predict an interpretation report, our first guess is likely to be that we used the wrong in formula (1.2). We postulate something different in the context. This is a legitimate move, but it must be tested severely. The astronomer must postulate a specific orbit for the unknown planet, show that it explains the anomaly, and seek independent confirmation of its existence--usually with a telescope. Similarly, the discourse analyst must specify the suspected context precisely, show that then predicts the interpretation report accurately, and seek independent confirmation of the crucial aspects of that context.

This last requirement brings up a good question however--how do we validate a hypothesized knowledge base? We will assume we have fixed upon a particular set of similar speakers, listeners, and global contexts, for which there is available a reliable source of interpretation reports. We have probably done this by picking an extensive corpus of texts produced by a single speaker or a coherent set of speakers, for which we may consider ourselves a part of the intended audience. A knowledge base, both its occasional and more permanent parts, and an interpretation procedure constitute a theory of the corpus, just as, in the large, physics and geology are theories of their own phenomena. The theory will generate interpretations that are formal objects, necessarily specified more precisely than can be verified by observables, but anchored in observable interpretation reports at various crucial points.

A theory of the corpus develops in a way analogous to the way in which, according to Lakatos (1970), a scientific research program develops. Lakatos distinguishes between the basic core of a theory--those principles that are exempt from counter-evidence--and the protective belt of auxilliary hypotheses that are adjusted to fit the evidence. In fact however, in a theory of a corpus, just as in a theory of natural phenomena, there is not just a two-fold division, but a set of layers. In discourse theory, when prediction fails, we are likely first to adjust our assumptions about the immediate context, then those about general knowledge. Only with greater reluctance do we modify the interpretation procedures. Nearly impervious to challenge are the overall theoretical framework and the computer metaphor itself. In general, we want to make those changes that are least consequential for the rest of our theory.^1.11 A sequence of such modifications yields a series of theories. A series of theories is progressive, in Lakatos' terms, if the successive theories cover more and more of the available evidence. In discouse, this would mean we were able to modify , and perhaps , in a way that correctly predicts more and more interpretation reports in more and more texts of the corpus. In validating a knowledge base and an interpretation procedure, a progressive series of theories of a corpus is the most we can hope for.

All of this makes discourse theory a difficult enterprise, but there was never any reason to suppose it would be an easy one. Nevertheless, as I hope to show in this book, a quite manageable and well-defined program of research is indicated. I will examine the nature of discourse theory in more detail in the final chapter of this book. In the meantime, the reader can view the book as laying the groundwork for such a theory of discourse, by defining precisely what , , , and indeed must look like.

Our primary data is therefore interpretation reports. But there are several commonly employed kinds of interpretation reports we will not use here. One is the reports that a subject gives of what he remembers of a text in a recall experiment. This is extremely complex behavior involving interpretation, memory, and creativity; there are many confounding factors; and there is no particular reason to believe such reports give us very direct views of memory structure, the interpretation processes, or anything else. Beaugrande (1980) gives an excellent critique of such a recall experiment.

Secondly, although it is assumed people can reliably report on their interpretations, it is decidedly not assumed they can reliably report on their interpretation processes. Introspection does not provide data. Introspection is a mode of accessing one's intuitions, and the role of intuition in discourse theory is just what it is in any scientific or critical enterprise: it is a source of hypotheses. It has nothing to do with the validity of what is hypothesized. We process language and we have an intuitive folk theory about how we process language, and there is no reason to believe they have very much in common. There are certainly occasions when we are quite conscious of part of the interpretation process; we may mull over the meanings of certain words at length. Much else seems to happen just at the edge of consciousness. Most of what occurs, however, is probably deeply unconscious, in the sense that it is in principle inaccessible to introspection. Belief and inference play a central role in discourse interpretation theory but only as theoretical constructs, not as data to be explained, not as something that can be introspected about and reported on.^1.12 Correspondingly, the intuitive plausibility of a hypothesis may cause a researcher to pursue the hypothesis with special vigor, and it may win other researchers over to an approach, but it has nothing to do with the validity of the hypothesis and does not in any way constitute an argument for or against a position.

More generally, a theory based on the computer metaphor is necessarily a ``deep'' theory, in the sense of Moravcsik (1980); the formal explanatory machinery is typically much greater in scope than the observable data it is intended to account for. The formal machinery must produce results that can be interpreted as corresponding to the observed data to be explained, but we cannot in general expect experimental confirmation of the details of the formal process that led to these results. For example, for sentence (1.1) it is proposed in Chapter 6 that we are able to resolve the definite noun phrase ``the index'' by drawing an inference based on an axiom in our knowledge base that says that many books have indexes. Nevertheless, we should not expect to find any direct evidence of this inference being drawn other than a reader's report of what he resolved ``the index'' to.

All of this does not mean, by the way, that no processing conclusions can be drawn. Consider, for example,

Who does ``he'' refer to? Most people will reply ``John''. But the complete text could have been

Hearing this text we sense that first we resolve ``he'' to John, but then as the second sentence proceeds, we change the resolution to Bill, and our report about (1.3) substantiates this. This seems to indicate that we have at least two strategies for resolving pronoun references--one independent of the semantic content and the other not. This is an example of drawing rather strong conclusions about processing from people's reports about interpretations, which are fairly accessible.

Finally, there is the exotic data. It is common in cognitive psychology to seek correspondences between reaction times and complexity properties of algorithms suggested by the formal theory. But I have ignored all data on timing of mental processes. Questions of timing and complexity are very much dependent on the architecture of the machine. We know very little about the architecture of the brain, but we do know that it is parallel. Present-day computability theory has not developed sufficiently advanced formalisms for parallel computation for us to attempt to build formal models of language processing that will explain these timing results. This is not an argument against conducting such experiments and using the computer metaphor in an informal way to draw the most general possible conclusions about processing from them. But where we take the requirement of computability seriously, I think we have no choice but to set these experimental results aside for future consideration.^1.13

 

It was said at the outset that this enterprise is a psychological one. We are now in a position to elaborate on this statement somewhat. The data that it seeks to account for--interpretation reports--is certainly psychological data. But how deep in the theory do the claims of psychological reality go? Am I claiming, for example, that anything like formal logical expressions actually exist in the brain?

Here I am fundamentally in agreement with Chomsky's position on ``psychological reality'' as expressed in Rules and Representations, pp. 106-112. He argues that regardless of the psychological phenomena we are seeking to explain, ``we should be willing to say at every stage that we are presenting psychological hypotheses and presenting conditions that the `inner mechanisms' are alleged to meet.'' But I would perhaps emphasize ``hypotheses'' over ``psychological''. I think it is necessary to hold all psychological (and other) theories at arm's length. The procedures and data structures of a psychological theory are usually referred to as ``cognitive processes'' and ``mental representations'', and the justification for this is usually that one should make the ontological assumptions that one's best theory indicates. This is not an issue of fact, however, but an issue of how we will talk about things. One can certainly never expect to find direct experimental evidence for all of one's theoretical constructs, in psychology or any other science. It will generally be more convenient to speak of representations of discourse and knowledge as being in the mind and to refer to the processes of interpretation as cognitive, without embedding such talk in hypotheticals and scare quotes, although I hope not to carry this to too detailed a level. But the reader should not lose sight of the fact that when this language is used, we are in the hypothetical world of explanation and not in the ``real'' world of data. Whether or not a theory is psychological does not depend on its ontological assumptions.

The phrase ``psychological reality'' is often used as a slogan for making one's theory responsible to particular kinds of exotic data. This frequently takes the form of a call for theories of ``performance'' rather than theories of ``competence'', in Chomsky's terms. Which is intended here? There are three aspects to the notion of ``competence''. As discussed above, the first is the assumption that grammaticality judgments and other interpretation reports are reliable, priveleged data. This is assumed here. It is this that makes prediction possible. The second is that such reports have something to do with the listener's interpretation procedures. This is also assumed here. Otherwise, a theory of interpretation reports would not be a theory of interpretation. The third, and least important, aspect is the idealization of the ``speaker-listener, in a completely homogeneous speech-community, who knows its language perfectly...'' (Chomsky, 1965). In this final sense, what is presented here is not a competence theory. Rather the interpretation procedures are explicitly made parametric on the individual listener's perhaps idiosyncratic knowledge base. It is a performance theory in that such conditions ``as memory limitations, distractions, shifts of attention and interest'' have to be built into the theory somehow in order to explain how different interpretations can result on different occasions. Moreover, in the case of syntax, in Chapter 4, what is essentially a competence grammar is developed, but it is then shown how its rules are deployed in time and can yield interpretations for sometimes quite incompetent productions.

The theory aims for psychological reality in that it seeks to explain some psychological data. It falls short of psychological reality in that it does not attempt to explain all psychological data. But in this is does not differ from other theories.