The Ascent of Babel: An Exploration of Language, Mind, and Understanding

by Gerry T. M. Altmann

  heft lemisphere [from left hemisphere] He dealt a blushing crow [from He dealt a crushing blow] a kite ream cone [from an ice cream cone] SM is souved [from Soup is served] Ouch! I have a stick neff [from Ouch! I have a stiff neck]

  These cases are straightforward exchanges of phonemes. And although Spooner was most famous for those of his errors which led to real words being produced (the `blushing crow' example is one of his), phoneme exchange errors often lead to the production of anomalous nonwords. Interestingly, when `an ice cream cone' became `a kice ream cone', the word that should have been `an' became `a', reflecting the fact that the following word now started with a consonant. So the process responsible for changing `a' into `an' must happen at a stage in the processing that takes place after the specific words have been selected, and after their phonemes have been selected.

  Finally, here is an error which illustrates one last thing that can go wrong once everything else, including the phonemes, has been correctly selected:

  Blear flue sky [from clear blue sky] pig and vat [from big and fat]

  Here, information about whether or not a phoneme should be voiced has exchanged (see Chapter 3). The phonemes /g/, /b/, and /v/ are produced with the vocal folds vibrating, and are voiced, whereas /k/, /p/, and /t/ are not, and are unvoiced. The first phoneme of `clear' should have been unvoiced, and the first phoneme of `blue' voiced. But the relevant information switched around, so that `clear' became `gear' and `blue' became `plue'. What has exchanged here is one of the component properties of each phoneme. Other properties also can exchange.

  All of these errors involved either an exchange of elements, or (in the case of the 'I disregard this as precise' error) the movement of a single element. But there are other errors too. Sometimes, the exchange is never completed, so that something from later on in the utterance can replace something earlier on, but without the corresponding movement of that earlier thing to later on (e.g. 'the pirst part' for `the first part'). And sometimes the earlier thing can replace a later thing, without that later thing moving to the earlier position (e.g. 'beef needle' for 'beef noodle').

  There are a number of fundamental things that all these errors, taken together, tell us about the production process. They provide evidence both of the nature of the elements that are queued, or buffered (anything that can move from later in an utterance to earlier is one such queuing element), and of the ordering of the different stages in the production process, with different kinds of error occurring at each different stage.

  There is one further piece of evidence that is fundamental in terms of shaping an account of how production proceeds: these errors are not just random events. If they were, then anything ought to be able to move to anywhere else. But there is one element which, surprisingly, is never involved in these kinds of error-the syllable. The sentence 'He plays the trombone' could never give rise to an error like `He troms the playbone' because `trom' is neither a word nor an affix. It is a syllable, and syllables never migrate. Parts of syllables do, but never whole syllables, unless they are also a word, or an affix. We shall come to this a little later on, because syllables do none the less have a vital role to play in the production process. But there is another, non-random feature of these errors which is even more important: there are no errors in which an entire word swaps with just a phoneme, or where the first phoneme of a syllable swaps with the last phoneme of a syllable, or where an affix at the beginning of a word (a prefix) moves to the end of another word. In all cases, an element can only exchange with an element of the same kind, or can only replace an element of the same kind, or can only move to a position normally occupied by an element of the same kind. Where does this like-for-like constraint come from?

  Sketching out the utterance

  When we speak, we provide our listeners with a specification (a builder's plan) of what to construct within their mental models.

  And like all specifications and plans, this one cannot come into being instantaneously. Instead, it must be created in stages, with something more being added to the plan at each stage. The fact that errors affect only one kind of element at a time suggests that the production of an utterance involves a sketching-out process in which one level of detail, involving just one kind of element, is sketched out before another. And as we have seen, different kinds of errors provide important clues regarding the sequence in which things are sketched out, and the nature of the elements that can be independently queued or buffered (that is, added to the sketch as an independent piece of information).

  To begin with, the utterance is sketched out as an ordered sequence of concepts. These concepts may then be fleshed out some more, so that information about the actual words that will express those concepts is added to the sketch. But this information is again only in sketch form. For instance, the concept corresponding to a verb like `slants' would be fleshed out to include the information that a word corresponding to the concept `slant' is required, and the information that it must be inflected for third person singular. The part of the sketch corresponding to the concept `slant' would be fleshed out to include information about the actual word that will be used to express it. Similarly for the inflection. This would include information about the onset of each syllable (the `sl' of `slants'), and the remainder ('ants'). And the remainder would be further sketched out in terms of its component phonemes: the vowel and the consonants following. Finally, these phonemes would be sketched out in terms of their component properties, including the ways in which they should be articulated, whether they should be voiced or not, and whether the mouth should close completely or not. These properties, or features, specify the motor program (the muscular movements) that causes the actual articulation.' _link_

  This sketching out process is nothing more than the operation of the neural circuitry. One way to think of it is in terms of different circuits corresponding to the different levels of the sketch, with the activation of one circuit causing the activation of the next, one after the other. Some circuits probably feed into several others which, operating in parallel, then together feed in to a single circuit further along the chain. But how could a single chain of circuits simultaneously encode the sketch for more than one concept? In other words, if this chain of circuits encodes the sketching out for what should become `slants', how can this same circuit also encode the sketch for a later word in the sentence-for what will become `writing'? How does the brain do queues?

  There are no real queues in the brain, just connections between neurons, and patterns of neural activity that happen when some of those neurons become activated and in turn activate the neurons they are connected to. A mental queue is nothing more than different patterns of activity, corresponding to the different things in the queue, with one pattern (the first thing in the queue) more active than another (the second thing) which is more active than another (the third) and so on. The equivalent of things getting removed from the queue is when that first pattern rapidly decreases in activity (having first triggered some neural consequence-another neural circuit), leaving what had previously been the second most active pattern to become more active, to the point where it also triggers some neural consequence, before also decreasing in activity.' _link_ It is as if there is something pointing at each pattern in turn and telling it to become more active as a function of how near the top of the mental queue it is. Exactly what controls this changing profile of activation is unclear, although in Chapter 13, on artificial brains, we shall see how the activation of a single set of (artificial) neurons can lead to a sequence just like this. What matters for the moment is that the brain can do queues.

  The basic idea, then, is that there is a chain of circuits in which each circuit encodes some particular level of the sketch. If a circuit at one end of the chain encodes concepts, say, and a circuit at the other end encodes phonemes, then the pattern of activity spreading across that entire chain will correspond to the sketching out of a particular concept in the manner outlined earlier. And there may be a seco
nd, less activated pattern also spreading across that entire chain, which corresponds to the sketching out of another concept in the queue. Within any one circuit, components of these two patterns will both exist, one more strongly than the other. In effect, each neural circuit encodes its own miniqueue.

  These different neural circuits do not need to be physically separate. It is possible, in principle, for the same set of neurons to encode different circuits. It is also possible for the same set of neurons to encode a queue in which the same phoneme, word, or whatever, appears at different positions within that same queue. Saying `bath the baby', for instance, requires the /b/ phoneme to appear at different positions within the phoneme queue. Some of the artificial neural circuits described in Chapter 13 do just this. For the moment, whether the different neural circuits are encoded across different neurons, or the same ones, does not matter.

  If each neural circuit encodes its own mini-queue, then it follows that if one pattern within that circuit gets a bit of a boost (we shall see why it might in a moment), it could swap positions with another pattern that is `higher up' that circuit's queue. And because each circuit represents just one kind of element, any misordering errors within a queue will affect only one kind of element-the like-for-like constraint.

  Why might a pattern of activity get an error-causing boost? In the first experimental study to elicit speech errors back in the early 1970s, Bernard Baars and colleagues at the University of California in Los Angeles asked people to read pairs of words as quickly as possible. They made errors, like `barn door' instead of `darn bore', as much as 15% of the time, but only when the pair was preceded by pairs such as `born dart' which shared the initial phonemes of the error they hoped to induce. So if they saw `born dart' then `darn bore', they were likely to say `born dart' correctly, then `barn door'. Presumably, the errors occurred because the pattern corresponding to the queue for syllableinitial phonemes (/b/ at the top of the queue, then /d/) had existed previously (in response to `born dart'), and some residual activity may still have been present. In support of this is the finding that errors are more likely the faster the speech rate, which gives even less time for those patterns of activity to reduce enough (tongue-twisters are no fun when said slowly). But not all errors involve overlaps with patterns that had existed previously, or with patterns later on in the queue, and no doubt some errors are simply due to random noise within the system.

  But is it not a little contrived to suggest that there is a queue devoted only to syllable-initial phonemes? It would certainly explain why they can be subject to errors and can swap places with one another. But is it reasonable to suppose that there is one queue for syllable-initial phonemes, another for syllable-final phonemes, and another even for the vowel in the middle? The answer to this question lies in the manner by which the neural circuitry itself develops. In Chapter 9, which dealt with neural circuitry and meaning, we saw that patterns of neural activity change as a function of experience; they come to reflect those experiences. Things which can occur in similar contexts will give rise to patterns of activity that are similar, but with subtle differences reflecting the actual differences between the different things. To take a concrete example: phonemes can occur in fairly similar contexts, although each is subtly different from the others. So all phonemes will give rise to similar, but subtly different, patterns. In effect, our experience of phonemes causes the development of a neural circuit-a set of connections between neurons which gives rise to the patterns of activation corresponding to the different phonemes. But phonemes at the beginnings of syllables necessarily occur in contexts which are different from those in which phonemes at the ends of syllables occur. This difference will give rise to different components in the neural patterns that encode those phonemes-and these different components constitute different circuits.

  Finally, where do syllables figure in all of this? Syllables are articulatory gestures. The movement of the articulators does not proceed phoneme-by-phoneme, but syllable-by-syllable. Try saying the phoneme /b/ on its own. You simply cannot. You can say /ba/, or /bi/, or /b some-other-vowel/. In other words, you can only say it in the context of a syllable. Vowels, unlike consonants, can be uttered as single phonemes, but then they can also be single syllables-the word `a' is a single phoneme, a single syllable, and a single word. Syllables, therefore, are the last thing to be added to the sketch. Everything that comes before, including phonemes and their features, serves this final act. Unlike the other components in speech production, syllables specify actions. The neural circuitry associated with these specifications is similar, in some respects, to the circuitry responsible for other kinds of muscular movement. And that is what makes syllables different-they are connected to a different kind of circuitry, and this is reflected in the finding that they are not subject to the same kinds of misordering as the other elements implicated in speech production.

  Beyond speech errors

  Psycholinguistics owes much to the imperfect mechanism it studies-if we made no errors, there would be no basis for speculating about the kinds of mechanism that could give rise to them. But there are other ways of studying that mechanism, although the issues remain the same. For example, having established that there are different kinds of mental queue, we could try to determine how long each queue was by looking at how far apart different kinds of element need to be before they can no longer exchange (words exchange over far greater distances than phonemes, for example).

  Much of the time when we speak we do nothing of the sort. We pause. But the pattern of pausing in normal spontaneous speech is quite different from the pattern when reading aloud. Reading aloud does not require any planning at the conceptual level (that is, at the level of the content of what is to be said); the words and their order are determined by the text. Basically, all the speaker need worry about i; getting the sounds right for each word-putting the information about those sounds onto the appropriate queues in the right order. But when is that information put onto the queues? And how far ahead does that information stretch? If the answer to the first question i~ `during the pauses', then the answer to the second would be `as far ahead as the next pause'. A big 'if, but it leads to some interesting conclusions.

  Marilyn Ford, an Australian psycholinguist working in the USA in the early 1980s, measured the brief pauses in a large sample of spontaneous speech. She found that the majority of the pauses divided each utterance into `chunks', where each chunk was a clause, containing just one verb. In effect, each of these clauses was a single who-didwhat-to-whom statement (see Chapter 7 for discussion of who-didwhat-to-whom). This suggests, if queuing does indeed happen during pausing, that the concepts associated with each who-did-what-towhom statement were put on the mental queue just before each such statement was to be uttered. Apparently, then, the concept queue is as long as necessary to convey the who, what, and whom.

  Around the same time that Marilyn Ford was measuring the pauses in spontaneous speech, Paul Gee and Francois Grosjean in Boston were doing the equivalent with read speech. They found, to simplify somewhat, that the pause pattern most naturally divided each sentence into chunks containing a single content word and any function words that preceded it. In effect, each chunk was whatever was required to express a single concept and its grammatical role. This suggests that the information regarding the sounds of each word was being queued more or less one concept (one who, what, or whom) at a time.

  In a sense this is quite counter-intuitive-surely we pause and hesitate far more in spontaneous speaking than during reading aloud? Yet it looks from these data as if there were more pauses (or at least, people paused after shorter intervals) in the reading case. In fact, during normal spontaneous speech, there are frequent pauses all over the place. But the major pauses, in Ford's measurements, occurred between clauses. In fact, pausing is not the only thing that divides speech up into chunks. The melody and rhythms of speech do too, and these also often reflect the conceptual and clausal structure. It is possible, even, that how
far each queue extends is in part determined by those melodies and rhythms. This is the focus of much research.

  Beyond the dysfluencies of speech

  Until the late 1980s, most of what we knew about how we speak was based on the dysfluencies with which we speak. But since then, research into speech production has undergone a small revolution, much of which has been inspired by a group in The Netherlands headed by Willem Levelt. Experimental techniques have been developed there (and subsequently elsewhere) which allow the researcher to delve more deeply into the individual stages of the process, and monitor what is going on as it happens. And yet, despite the major advances that this offers, the basic model of speech production that is almost universally accepted within the psycholinguistic community is the same one as had been developed from the error data. What has changed is that we can now both confirm that model, and add to it.

  Many of these experimental techniques involve a variation on the priming theme. In Chapter 6 we saw that the time it takes to recognize a target word can be shortened if a word is presented beforehand which is related in meaning (as `brush' is to `broom'). It turns out that the time to say a word can also be reduced if an appropriate prime is presented beforehand. How the target words are presented varies from study to study. In many of these studies, people simply have to name a picture of a common object. But whether or not the time to name that picture can be reduced, by presenting a prime beforehand, depends on a number of factors.