Edward Sapir begins his book Language with these two sentences:
Speech is so familiar a feature of daily life that we rarely pause to define it. It seems as natural to man as walking, and only less so than breathing.
We read these sentences without difficulty, and we would hear them with similar ease. Yet when examined more closely, this text poses a wealth of problems.
Consider the pronoun ``it'' at the end of the first sentence. How do we know it refers to speech rather than daily life? We cannot tell just from the syntactic structure, for the following sentence has the same syntactic structure, and we resolve the pronoun the other way:
Secrecy is so pervasive a feature of our foreign policy that it is hard to learn anything about it.
No simple view of semantics helps us here either. It is just as possible to define daily life as it is to define speech. The answer apparently involves knowing a complex relationship between the familiarity of something and our lack of awareness of it. But how do we access this knowledge and use it to resolve the pronoun reference?
Or consider the word ``of''. ``Of'' can mean many things - possession of a physical object, as in ``a book of John's''; identity, as in ``the city of Boston''; participation in an event, as in ``the destruction of Rome''; and other more complex relations. What does it mean here? A feature must be a feature of something, and daily life is that something; this is what ``of'' conveys. But how do we know this? The word ``feature'' alone does not give us the answer, for one can say,
Speech has the feature of familiarity,
in which ``of'' is the ``of'' of identity.
Sapir presupposes that the reader knows speech is a feature of daily life, but what kind of relationship has to hold between two things, such as speech and daily life, for one to be a feature of the other? The phrase ``so familiar a feature of daily life that...'' makes a comparison along a dimension defined by the words ``familiar a feature of daily life''. What is the specific nature of that dimension, and how are we able to locate a place for speech on it. If we did frequently pause to define something, what would that say about where it lay on this dimension?
This by no means exhausts the problems presented by the first sentence. The word ``speech'', for example, is lexically ambiguous between language and a formal address that someone delivers. The syntax of the phrase ``so familiar a feature of daily life that...'' is quite unusual. We might analyze it as an adjective followed by a noun phrase, but not just any noun phrase will do. We can't say, for example, ``so familiar some feature of daily life that....'' What is the general rule? By ``we'' does Sapir mean just himself and his reader? The set of all linguists? The set of all people? Does it matter whom he is referring to, and if it doesn't, how do we know it doesn't? For there are many times when it does matter. What is the implicit relation between day and life that is encoded in the phrase ``daily life''? We can ask about ``rarely'' how many times counts as rarely and what is ``pausing to define something'' that it can occur rarely. What is speech that it can be defined? In what sense, if any, does pausing enable us to define speech?
The second sentence presents the same wealth of problems: How do we know ``it'' again refers to ``speech''? What are speech, walking and breathing that they can be natural? How does the phrase ``to man'' alter the meaning of ``natural''? What is the nature of the dimension ``naturalness to man'' that allows us to compare speech, walking and breathing on it? How are we able to construct this dimension for comparison, seemingly so effortlessly. What does Sapir mean by ``only'' - that breathing is the unique thing that is more natural to man than speech, or does ``only'' mean something more like ``slightly'' or ``just''? What predicate is the word ``so'' standing in for? What function does the word ``seems'' play, and why was it chosen rather than ``is''?
It is not just the words and phrases in single sentences that present difficulties. The very fact that two sentences occur next to each other in a text, and sound right together, requires explanation. What is the relationship between Sapir's two sentences that we are not surprised that they are part of the same text? Intuitively, we might say that they elaborate on the same theme. Pressed to be more specific, we might say that both sentences are making comparative statements about the familiarity or naturalness of speech. How do we see past the words used to this similarity of expressed thought, and why does this similarity seem to confer a feel of coherence to the text?
Finally we might ask why Sapir is telling us this? Why did he write these two sentences? It is not new information. Indeed, it is clear he expects us to agree with what he says. But as soon as we agree with it, it suddenly becomes less true. His writing it and our reading it have suddenly made speech seem less familiar and less natural, and this appears to be Sapir's goal--to make speech seem suddenly problematic. The next sentence in the text is ``Yet it needs but a moment's reflection to convince us that this naturalness in speech is but an illusory feeling.'' The title of the first chapter is ``Introductory: Language Defined'', and the purpose of the entire book Language is to subject speech to close scrutiny. In short, Sapir wrote his first two sentences for very much the same reasons that I wrote my first paragraph--to set up the first half of a contrast, the second half of which makes an investigation, and a book, possible.
I chose the Sapir text for a number of reasons--its ecological validity, its good but nonliterary style, the prestige of its author, the possibilities of self-reference--but not for the discourse problems it posed. It is not at all unusual in this regard. Virtually any two consecutive and moderately complex sentences would have served as well. With all of the difficulties presented by ordinary texts, how is it that we understand them so readily? More seriously, how is it that we understand them at all?
From one point of view, the point of view of daily life, all of these problems are not really problems. How do we solve them? We just do. But if we ask about the mechanisms that underlie our ability to solve these problems, and we are very strict about what we mean by ``mechanism'', all of these become significant problems indeed.
For a number of years, researchers in the fields of artificial intelligence (AI) and cognitive psychology have looked at discourse from just such a point of view. The theory developed in this book arises out of this AI tradition. Workers in AI have tried to build natural language understanding systems, and the computer has placed quite stringent limits on what counts as a mechanism. This work is of more than just technological relevance. Today, only people understand language. Efforts to get computers to understand language must therefore draw on what is known about how people do it. Conversely, in spite of the substantial differences in architecture and power between the human brain and the present-day computer, there is embedded within most natural language processing systems a theory, implicit or explicit, of how language would be comprehended by any intelligent entity, hence by humans. If a theory of discourse is framed at a sufficiently abstract level, it may apply equally well to human and computer understanding.
The fundamental lesson so far of this AI research is that we understand discourse so well because we know so much. We understood the Sapir text because we know a lot about speech, familiarity, features, daily life, and so on. But we do not just have the knowledge, we are able to use the knowledge to make sense of the text. The central problems in understanding how people interpret discourse are therefore how to characterize the knowledge that people have and the processes they use to deploy that knowledge in the task of interpreting discourse. It is the aim of this book to address these two problems.
Before going on, it will be useful to fix certain crucial parts of our vocabulary. As a start, ``discourse'' will be taken to mean people communicating with each other, although in the next section we will focus in more on our immediate concerns. A ``text'' is a fragment of discourse produced by a single speaker. A fragment of discourse in which more than one participant speaks is referred to as a ``dialog''. The words ``speaker'' and ``listener'' will often be taken to cover writers and readers as well. The word ``listener'' is used rather than the more common ``hearer'' because listening is a more active process than hearing. The mapping from some representation in a speaker's mind into a text will be referred to as ``production'' rather than ``generation'', since the latter has acquired a very specific technical meaning in linguistics. The mapping from the text into a representation in the listener's mind will be referred to as ``interpretation'' rather than ``comprehension'', since toward the end of the book I wish to divorce the process from individual cognition.