4 CompreShensBoos understanding what we hear and read Understanding language, like producing it, is such an automatic task that it may appear to be a relatively straightforward process. Sounds or letters strike our ears or eyes in a swift and linear fashion creating words, which in turn very quickly form phrases, clauses, and sentences so that comprehension seems to be nothing more than the recognition of a sequential string of linguistic symbols, albeit at a very rapid pace. What appears on the surface to be linguistically transparent, however, turns out to be almost impenetrably complex from the perspective of psycholinguistics. What is apparent from the vast research into the comprehension of spoken and written language is that people do not process linguistic information in a neat, linear fashion; they do not move smoothly from one linguistic level to another as if they were riding a lift that began on the ground floor of phonology and finally stopped at the top floor of meaning. The research shows that in most situations, listeners and readers use a great deal of information other than the actual language being produced to help them decipher the linguistic symbols they hear or see. The comprehension of sounds Here is a simple example of how what we hear is influenced by psycholinguistic variables and is not just the accurate perception of the sequences of sounds or words that hit our ears. In one psycholinguistic experiment, a set of sentences was played to a group of listeners who were asked to write down the sixth word in each of the following sentences. (i) It was found that the _eel was on the axle. {2.) It was found that the _ecl was on the shoe. (3) It was found that the _eel was on the orange. (4) It was found that the _eel was on the table. Notice that in every case, the subjects heard eel as the key word in the sentence, but most of the subjects claimed they had heard a different word for each example- specifically, wheel for (1), heel for (2.), peel for (3), and weal for (4). The insertion of a different missing sound (phoneme) to create a separate but appropriate 'eel word in each sentence is called the phoneme restoration effect. Under these conditions, listeners do not accurately record what they hear; rather, they report what they expected to hear from the context, even if it means they must add a sound that was never actually spoken at the beginning of the target word. Several simple but significant observations can be drawn from this sample of the early psycholinguistic research into the nature of comprehension. First of all, as just illustrated, people don't necessarily hear each of the words spoken to them. Comprehension is not the passive recording of whatever is heard or seen; listeners are not tape recorders nor readers video cameras. Second, comprehension is strongly influenced by even the slightest of changes in discourse which the listener is attending to. In these examples, except for the last word, each of these sentences is identical. Finally, comprehension is not a simple item-by-item analysis of words in a linear sequence. We don't read or hear the same way we count digits sequentially from one to ten. Listeners and readers process chunks of information and sometimes wait to make decisions on what is comprehended until much later in the sequence. It is the last—not the sixth or 'target'—word in each of the four examples above which dictated what the listeners in the experiment reported they heard. We don't seem to listen to each word individually and comprehend its meaning in isolation; we seek contextual consistency and plausibility, even if it comes to adding a sound or inventing a word that wasn't actually spoken. This chapter then reviews some of the ways in which psycholinguistic processes affect the way listeners and readers comprehend language. r Although, in the course of everyday conversation, we don't hear voweis and consonants as isolated sounds, we can, with the help of machines, measure acoustic information extremely precisely. The /p/ in the following English words is pronounced slightly differently depending on where it occurs in the word or what other sounds follow it. The initial /p/ of 'pool' is pronounced with puckered lips but the 'same' /p/ in 'peel' is spoken with the lips spread, and neither of these /p/'s sound quite like the /p/ in 'spring'. Although these details may seem trivial to a native speaker of English, they are significant enough acoustically to be heard as contrasting phonemes in other languages. Despite these differences, and other variations of /p/ that could be cited in countless other examples, native speakers of English claim they hear and pronounce the same /p/ sound. Notice that for these and most of the other examples, we spell the sound with the letter 'p' and furthermore, despite all the variations in /p/, native speakers of English almost never confuse any manifestation of the /p/ sound with /b/, which is acoustically very similar. Recall that in the discussion of the articulation stage in Chapter 3, we saw that there was a sizeable phonetic difference between the initial /k/s of 'keep' and 'kid' and the /k/ sound which begins the word 'cool'. Phoneticians have been fairly successful in writing rules that predict which precise acoustic form of /p/ is pronounced (or heard) under which phonetic condition; nevertheless, they have been unable to explain how this variation is processed by the mind or how all the phonetic differences which occur among all the many languages of the world can be accounted for in terms of the common, universal processes of perception that are shared by all humans. Although the exact details of this acoustic processing have yet to be resolved, psycholinguists have come up with some explanations for this most fundamental level of comprehension. Suppose we are engaged in conversation with a friend and are discussing two other acquaintances with similar sounding names—'Benny' and 'Penny'. What phonetic information do we employ as we listen to distinguish these names which are identical in pronunciation except for the initial consonant? Phoneticians have discovered that the main feature which English speakers attend to, albeit unconsciously, is the Voice Onset Timing (VOT) of the initial consonant. Using instruments which are sensitive enough to measure contrasts as small as milliseconds in the duration of speech sounds, they have demonstrated that the most significant acoustic different between English consonants like Ihl and /p/ is the length of time it takes between the initial puff of air that begins these sounds, and the onset of voicing in the throat that initiates any vowel sound which follows the consonants. Since almost all the other phonetic features of this consonantal pair are identical, the crucial clue that separates the voiced Ihl and its voiceless counterpart /p/ is a VOT of a scant 50 milliseconds. This means that the correct comprehension of the name 'Penny', as opposed to the mistaken recognition of the similar sounding 'Benny', depends on an ability to perceive a voicing delay of one-twentieth of a second! The simple task of recognizing which person is being referred to during a conversation is based on your ability to isolate one subtle phonological feature from the myriad sounds hitting your ear and to make a split-second judgment. How do speakers of English, or any language for that matter, make these incredibly difficult decisions about speech so rapidly and so accurately? It appears that the acquisition of this phonetic ability cannot be completely explained only by exposure to, or instruction in, the language. In other words, native speakers do not acquire all of this acoustic information from direct experience with language, and as we learned in Chapter 2, parents and caretakers do not provide explicit instruction on these matters. Even phoneticians do not subject their children to hours of nursery training listening to minimal pairs like 'pie' versus 'buy'. Psycholinguists have discovered through careful experimentation that humans are actually born with the ability to focus in on VOT differences in the speech sounds they hear, and they have proven that rather than perceiving VOT contrasts as a continuum, people tend to categorize these minute phonetic differences in a non-continual, binary fashion. All of this has been decisively documented in experiments where native speakers of English listened to artificially created consonant sounds with gradually lengthening VOTs and were asked to judge whether the syllables they heard began with a voiced consonant (like Ihl which has a short VOT) and a voiceless one (like Ipl which, as was just pointed out, has a VOT lag of about 50 milliseconds). When subjects heard sounds with a VOT of about 25 milliseconds, about halfway between a Ibl and a /p/, they rarely judged the sound to be 50% voiceless and 50% voiced, they classified it as one sound or the other. This phenomenon is called categorical perception. Psycholinguists have been able to prove the presence of categorical perception in very young infants, through a series of cleverly designed experiments. And in equally ingenious research with several species of animals, they have found, by and large, that this kind of all-or-nothing acoustic perception does not exist in other species. Categorical perception is seemingly unique to human beings, and appears to qualify as one aspect of universal grammar (UG), the genetic propensity for comprehending and producing language which most psycholinguists believe is a uniquely human endowment. These experiments with VOT perception in human infants are one of the few solid pieces of evidence we have that UG exists and that at least part of human language is modular—that is, some parts of language reside in the mind or brain as an independent system or module. Although categorical perception of VOT is an ability children are born with, it is also influenced by the linguistic environment a child is raised in. Here lies the second part of the puzzle of how native speakers of English grow up with the intrinsic ability to distinguish instantly between the names 'Benny' and 'Penny'. Because the English language divides the VOT spectrum into two sets of sounds, for example the voiced and voiceless pairs of consonants Ibl versus /p/, /d/ versus /t/, and /g/ versus /k/, children learning English acquire the ability to use their innately specified gift of categorical perception to divide the VOT continuum into two equal halves, corresponding to the voiced and voiceless consonants just exemplified. On the other hand, children exposed to a different language, say Thai, which has three, not two, VOT consonantal contrasts, grow up after years of exposure with the ability to make a three-way categorical split. Thus Thai children rapidly acquire the ability to hear an extremely short VOT as Ibl (as in /bai/, the Thai word for 'leaf'), a slightly longer VOT as /p/ (a sound like the Ipl in the English word 'spring', as in /pai/, the Thai word for 'go'), and any VOT longer than 50 milliseconds as an aspirated /ph/ (a sound very close to the English Ipl and which is used in the Thai word, /phai/, which means 'paddle'). "When any language learner, whether a child learning their first language, or an adult a second language, is exposed to the VOT settings of a particular language over an extended period of time with lots of opportunities for acoustic input, it appears that they use their innate ability to hear speech sounds categorically to acquire the appropriate VOT settings. The successful comprehension of speech sounds is, therefore, a combination of the innate ability to recognize fine distinctions between speech sounds which all humans appear to possess, along with the ability all learners have to adjust their acoustic categories to the parameters of the language, or languages, they have been immersed in. We see then that learning to comprehend, like all aspects of language acquisition, is again a merger of both nature and nurture. The comprehension of words Sounds represent only a tiny and rather primitive component of comprehension. What about our comprehension of words? What psycholinguistic mechanisms affect lexical processing? Obviously, the comprehension of words is much more complex than the processing of phonemes. Because even short, one-syllable words are made up of at least several sounds, because these sounds may be written in different and inconsistent ways in various languages, because there are literally tens of thousands of words in the vocabulary of any language (in contrast to a few score phonemes), and, most importantly of all, because they convey meanings, the comprehension of words is indeed a very complex psycholinguistic process. One model that psycholinguists have adopted to account for this complexity is Parallel Distributed Processing (PDP). Using a model of cognition developed from recent research in neurology, computer science, and psychology, the PDP perspective argues that we use several separate but simultaneous and parallel processes when we try to understand spoken or written language. These processes are used at all levels of linguistic analysis, but play a particularly conspicuous role in the comprehension of words and sentences. One explanation, based on this approach, r for how we access the words stored in our mental lexicon is the logogen model of comprehension. When you hear a word in a conversation or see it on the printed page (as you have just done with this new term), you stimulare an individual logogen, or lexical detection device, for that word. Logogens can be likened to individual neurons in a gigantic neuronal network; if they are activated, or 'fired', they work in parallel and in concert with many other logogens (or nerve cells) to create comprehension. High-frequency words (like the word 'word') are represented by logogens with hair triggers; they are rapidly and frequently fired. Low-frequency words (like the word 'logogen' itself) have very high thresholds of activation and take longer to be incorporated into a system of understanding. By adopting this model, psycholinguists can account for the comprehension of words in several different ways: in terms of their spelling (for example homophones like 'threw* and 'through', which are spelled differently but pronounced alike); on the basis of their pronunciation (for example homographs like the verb 'lead' and the noun 'lead', which are spelled alike but pronounced differently), or in terms of the grammatical functions that the word might fill (for example 'smell' can function as either a noun or a verb, but 'hear' functions only as a verb and requires derivational or lexical changes, like 'hearing' or 'sound' when used as a noun). Finally, comprehension can be linked via PDP to the network of associations that are triggered by a word's meaning (for example the word 'leaf can very rapidly evoke images of trees, pages in a book, or even words which sound similar such as the verb 'leave'). A clear example of the usefulness of a PDP approach to the comprehension of words is an experience many of us encounter on an almost daily basis, what psychologists term the Tip-Of-the-Tongue (TOT) phenomenon. Because our long-term memory storage is better for recognition than for recall, we often know that we know a word so that, even when we can't recall it from our memory, it is on the tip of our tongue, and we can instantly recognize the word when it is presented to us. Psycholinguists have studied this frequent linguistic experience and have discovered several intriguing aspects of the TOT phenomenon. For one thing, the momentarily lost word isn't always completely forgot- ten; parts of the word are often subject to recall and, most commonly, these remembered fragments are the first letters or the first syllable. Suppose you are trying to recall an obscure word, say the word which refers to the belief that everything that happens to us has already been ordained by God. If we have actually acquired this word at some time in our life, then we usually have some TOT memory of how it begins. We think that it is a polysyllabic word which begins with 'pre-'. In trying to produce the word, we somewhat frustratedly experience the TOT phenomenon because we know we know the word, and we remember something about the term, but we simply cannot recall it on demand. Another intriguing aspect about examples like this is that although we cannot reproduce the word, we can instantly recognize any words that are not the one we are trying to recall. As soon as we see or hear the following words, we know that none of them is the TOT item we are searching for. prestidigitation pretension Presbyterian predilection We know that even though none of these word fits our ephemeral image of the target of our lexical search, they are all pretty good matches. For one thing, the TOT word we are trying to recall seems to end with an '-ion' like the set of terms above. Often we have vague memories of the beginning and the ending of TOT terms but not the middle, which is, so to speak, submerged. This so-called bathtub effect allows us to search for words in a dictionary, since memory of the beginning of the missing word allows us to access alphabetical files, and conversely, the memory of how the word ends allows us to use rhyming as one strategy to confirm whether the word we are searching for is among those on the page in front of us. Often, it is through these search strategies that we suddenly come up with the word, or recognize it insrantly if it is presented to us. At one moment we have only partial recollection, and at the next we remember the word is 'predestination'. Notions like logogens and PDP seem to be useful in explaining the TOT phenomenon under discussion here. We were able to recall, at least in this TOT example, the first and the last part of the logogen for the target word, and this allowed us to compare and contrast other words with similar logogens. Notice, too, that we are not confined to one type of comprehension or recognition processing when we are contrasting a TOT word with other possible targets. While we are looking at these words morphologically, we are also making other judgments. In the example above, 'Presbyterian' was rejected because of the capital letter (the word we were trying to retrieve was not a proper noun), and yet at the same time, we might have been dimly aware of a possible semantic connection since this word and our TOT term have something to do with theology while the other words on the list do not. Here our schematic knowledge, based on all of our life experiences, assists the lexical search process. A PDP model of comprehension is able to explain the very rapid and accurate way people make judgments about which, if any, words on a list are a temporarily forgotten TOT word because it accounts for the concurrent use of more general, or top-down, semantic information as well as more detailed, or bottom-up, 'bathtub' knowledge about the phonology or the exact spelling of the item being searched for. This example also demonstrates the effects of spreading activation networks. When you first try to recall a TOT word, it seems as if your memory is a complete blank and you have absolutely no clues about the word in question. Nevertheless, the more you think about the missing term and the more you contrast it with similar but not identical words, the more pieces of knowledge you activate so that the network of associations spreads. The first two items on the list, 'prestidigitation', and 'pretension', do not fit the lexical network you have established, but 'Presbyterian' does and, depending on your linguistic and schematic knowledge, even though this third word isn't a match, it helps accelerate the activation of lexical relationships so that eventually the target word you are searching for is reached. Lexical recognition and comprehension then are much more difficult processes to understand than the recognition of phonemes, but we have learned several things about these processes over the past few decades. First, we know that words are not stored solely in alphabetical order in mental 'dictionaries', although the bathtub effect demonstrates that this type of serial-order recognition and retrieval is available to us. We also store words according to how their last syllables rhyme, for example. Second, we have learned that comprehension is not an absolute state where language users either fully comprehend or are left completely in the dark. Rather, it seems that comprehension involves a dynamic, growing, and active process of searching for relevant relationships in spreading activation networks. The logogen model suggests that familiar words connect rapidly with other nodes in the network; unfamiliar words take time because the connections have not been automated. Finally, we see that people do nor rely on one general strategy to comprehend words, but simultaneously use both top-down information involving context and meaning and bottom-up data about the pronunciation and spelling of words to assist them in decoding the words they hear or read. From all this, it is manifest that listening and reading are not 'simple' or 'passive' activities. They require just as much complex and active mental processing as their more physically overt linguistic counterparts, speaking and writing. The comprehension of sentences But comprehension involves much more than the decoding of sounds, letters, and lexical meanings; it also involves the untangling of the semantics of sentences. Psycholinguists first began to examine the comprehension of sentences by basing their research on the model of sentence grammar originally proposed by Chomsky in the 1950s. Chomsky's model claimed that ail sentences were 'generated' from a phrase structure skeleton which was then fleshed out into everyday utterances by a series of transformational rules (hence the term Transformational-Generative (TG) grammar). In the original version of grammar, these transformations were plenteous and powerful, and they could create many varieties of 'surface structures' by rearranging, deleting, adding, or substituting words which were found in the 'deep structure' of the original PS skeleton. Using this model, psycholinguists immediately became interested in comparing the number of transformations used to derive sentences and the relative difficulty native speakers experienced in comprehending them. They based these early experiments on sentence pairs like the following. (1) The dog is chasing the cat. (2) Isn't the cat being chased by the dog? From the standpoint of TG grammar, {2) is much more complex than (i), not simply because it contains two more words ('n't1 and 'by'), but because unlike (1) which corresponds with the underlying PS sentence, {2) has undergone three transformational changes; it has been transformed into a negative, passive, interrogative sentence. Accepting this linguistic analysis for the moment, it is easy to see why psycholinguists thought that pairs of sentences like these might offer insights into the comprehension process. It seems logical that simple 'kernel' sentences like (1) are easier to comprehend and remember than complex sentences like {2). Psycholinguists who first experimented with this hypothesis called it the Derivational Theory of Complexity (DTC), because difficulty in comprehension was derived from the number of transformations that were added on to the original phrase structure of the kernel sentence. Several creative experiments were devised in the 1960s to test the DTC. For example, subjects were given a random assortment of sentences like the following and were then asked to recall both the sentence they had just heard and a string of words spoken immediately after the sentence. (3) The dog is chasing the cat. (4) The dog isn't chasing the cat. bus/green/chair/ apple/etc. car/blue/sofa/ pear/etc. (5) Is the cat being chased by the dog? bike/pink/table/ peach/etc. (6) Isn't the cat being chased by train/yellow/stoo!/ the dog? grape/etc. Researchers hypothesized that since working memory constrains the amount of new linguistic information we hear, and because each sentence got more and more complicated in a very quantifiable way, the subjects would remember fewer and fewer words following each sentence. That is, based on the DTC, it was claimed that (4) was (3) plus an additional transformation (the negative), (5) was (3) plus two additional transformations (the passive and the interrogative), and (6) was (3) with three additional transformations (the negative, the passive, and the interrogative). Accordingly, it was hypothesized that for sentences like (3), subjects would remember several of the words following the initial sentence (perhaps around six), but that for each successive sentence, subjects would remember one fewer word on the list because their working memory would be taxed by additional transformations. Initial experiments like this based on the DTC showed a very broad confirmation of the hypothesis—the number of words remembered at the end of each sentence seemed to correlate inversely with the number of transformations presumably required to generate each of the sample sentences. Nevertheless, even those original DTC experiments which appeared to support the basic hypothesis, contained within them evidence that all was not well. One disturbing result was that although in general the more transformations a sentence contained, the more difficult it was to process in an experimental situation, there were unexplainable exceptions to this generalization. In one experiment, which simply asked subjects to match kernel sentences listed in a column on the left with their transformed variants in a right-hand column, although the matching took longer for sentences that contained more transformations, there were several exceptions. Sentences like {4) in the examples on the previous page, which had undergone only one transformational change, into the negative, often took just as much time to match as sentences with two or three transformations, as in examples (5) and (6). This seems very odd given that the interrogative and passive transformations, all contained in (6), are much more complicated than the negative rule, which simply adds 'not1 after the verb 'is'. The passive rule, on the other hand, is actually a collection of cumbersome transformations with words being added and rearranged in a complicated way. This would lead one to suspect that, all things being equal, sentences in the passive would take much longer to match (or cause greater constraints on memory) than negative sentences, which have undergone only minimal syntactic change. These early warning signals that the DTC was not as straightforward or as insightful as was originally hoped led to further experimentation. Failure to replicate the apparent successes of the early research led to the demise of the DTC by the end of the decade. One such replication attempted to repeat the study which looked at the reputed effect of the DTC on memory for lists of words, except rather than have the subjects hear the list of words after they heard the test sentence, the subjects listened to the words before hearing the sentence. When this slight change in protocol was introduced, it was discovered that the grammatical form of each sentence had no effect on the number of words recalled, suggesting that a larger number of transformations in a sentence does not necessarily occupy more space in working memory. Another series of experiments demonstrated that semantics, rather than syntax, seemed to be the main determinant of comprehension difficulty. Thus, in the examples below, passive sentences like (7) took less time to process than active sentences like (8) because they were semanticaliy more plausible. In fact, sentences like (8) tended to be remembered as (7) because subjects retained a rough memory of the word order but were reluctant to admit that they heard the highly implausible event described by (8), even though it was an active sentence. Instead, they were quick to claim what they actually heard was the reputedly more 'difficult' but vastly more plausible passive counterpart (7). (7) The struggling swimmer was rescued by the lifeguard. {8) The struggling swimmer rescued the lifeguard. Another way in which semantics seemed to intervene as a more important variable than the DTC was the manner in which any negative sentence seemed to confuse the subjects. As has been already pointed out, the negative in English is a relatively easy syntactic rule, especially in the case of sentences which include the auxiliary 'be': 'not' (or its contracted form, 'n't') is placed right after the verb 'be' turning an affirmative sentence like (7} into its negative equivalent. 'The struggling swimmer wasn't rescued by the lifeguard.' Contrast this solitary grammatical change with the four operations that are needed to change an active sentence like (y) into its passive equivalent (10): first the subject and object are reversed; second the preposition 'by' is inserted before the original subject; r third the verb 'be' is introduced in the correct tense; and fourth the main verb is converted into its past participle form. (9) The puppy hid the bone. (10} The bone was hidden by the puppy. Despite all of the changes passivization entails, it still seems that negative sentences take more time to comprehend and are more difficult to remember than passives, a finding that further undermines the hypothesis that the DTC plays a significant role in comprehension. This initial finding led to some further psycholinguists inquiry into the innate difficulty of negatives and showed that negation, especially double or triple negation, is exceedingly difficult to comprehend, despite the fact that grammatically, it is a simple structure in English. As a simple but revealing confirmation of this finding, quickly try to work out whether the following sentence is true or false. (11) It's not true that Wednesday never comes after a day that's not Tuesday. Finally, the original linguistic model which psycholinguists had based their early experiments on had already undergone major revisions. Ironically, even as psycholinguists began their series of DTC experiments in the 1960s, Chomsky had already made extensive changes to the primitive version of TG grammar upon which the DTC studies were based. By the time they were being conducted, he had introduced a revision which reduced the number and the power of transformational rules and, concurrently, featured a more prominent role for semantics in his model of grammar. Not surprisingly, he has continued to introduce further revisions in his model so that today there are virtually no transformations at all. Now, some thirty years later, it is generally accepted that transformational rules, especially as they were conceived of several decades ago, are not psycholinguistically relevant. If transformational complexity does not affect comprehension, what does? Thanks to further experimentation on a wide range of variables, it seems, quite a few factors. For one thing, ambiguity seems to slow down comprehension time, as has been demonstrated by several studies that use phoneme monitoring tasks. These r sound like some sore of phonological measure, but they are in fact a method psycholinguists use to tap into the process of sentence comprehension. Subjects listen to sentences like the pairs below and are asked to press a button as soon as they hear a Ibl sound. This allows the experimenter to measure the subjects' reaction times between the moment they heard the Ibl and the instant they reacted. The underlying assumption is that sentences which contain more complex information in the clause preceding the target phoneme will create a correspondingly greater lag in reaction time. Notice that in sentences (12) and (14}, the words immediately preceding the target sound Ibl are ambiguous; men can 'drill' by using an instrument or by rehearsing marching formations, and 'straw' can refer either to dried grass or to a tube used for sipping liquids. In contrast, sentences (13) and (15) do not use ambiguous words immediately before the target phoneme. (12) The men started to drill before they were ordered to do so. (13) The men started to march before they were ordered to do so. (14) The merchant put his straw beside the machine. (15) The merchant put his oats beside the machine. Sure enough, the subjects took several tens of milliseconds longer to hit the button when they heard the Ibl for sentences (12) and (14), most probably because of the ambiguity of the words 'drill' and 'straw', which immediately preceded the target sounds. Subjects were significantly faster in responding to the Ibl for sentences (13) and (15), presumably because they did not have to process two different meanings for 'march' and 'oats', the words which they heard just before the target phoneme. Sample sentences like the ones just cited reveal another important finding in the psycholinguists investigation of comprehension. Just as words tend to be processed, at least in part, in a linear, 'left-to-right' order, sentences also seem to be understood sequentially so that each new word serves to add to the meaning of the words which immediately preceded it and, at the same time, helps the listener or the reader to anticipate the next word or words which will follow. This form of 'spreading activation' at the sentence level has led some psycholinguists to posit Automated 64 SURVEY Transition Networks (ATNs) which can be used to predict the next word or word sequence at any juncture of a sentence as it is spoken or printed. Attempts to program computers to make predictions using ATNs have met with limited success, and this particular approach is not very popular, largely because it seems too simplistic to explain sentence comprehension on the basis of the single process of sequential prediction. The Parallel Distributed Processing (PDP) model is a more robust alternative, because it suggests the existence of multiple and parallel sequences of psycholinguistic processes, operating concurrently whenever we attempt to understand novel utterances. But the general tendency for all listeners and readers to make increasingly confident predictions about the meaning of a sentence as it progresses is well-attested in psycholinguistics and is colorfully called garden-pathing. One well-documented example of this phenomenon is the way comprehension is temporarily impeded when the listener or reader meanders down the wrong garden path in comprehending a string of words. Consider the following utterance and imagine that you were listening to it for the first time. After the word 'mile', what word or phrase would you expect to hear? {16) Since Jay always jogs a mile seems like a short distance to him. You probably expected something like 'he' as in 'he is in fairly good shape' rather than the verb 'seems', and this probably temporarily confused you. In fact, you may have scanned the sentence once or twice to see if there was some misprint. Actually, if we adhered to the rules of English punctuation, there should have been a comma after the first verb, 'jogs'. Researchers asked subjects to read sentences like (16) while they measured their rapid eye movements scanning the text using specially designed contact lenses. They found that when the subjects saw the word 'seems', the eye-fixation time was much longer than at any other point in the sentence. On the other hand, when sentences like (16) were slightly modified so as to remove any ambiguity about the direction of the sentence, as in (17), the subjects did not hesitate after the word 'mile'. COMPREHENSION 6s r (17) Since Jay always jogs a mile this seems like a short distance to him. Again we see evidence that the linguistic structure of the sentence affects the processing rime. When our guesses about which direction a sentence will go are correct, as is normally the case if we know a language well, our comprehension is rapid, but if we choose the wrong path, as is easy to do in a specially designed sentence like (16), our comprehension is disrupted. And what slows us down is the way we expect words to fit together syntactically. Since 'mile' follows 'jogs', it is natural to assume that all the words ro this point match the Noun + Verb + Noun pattern we expect for simple sentences and combine to create the initial clause, 'Since Jay always jogs a mile'. But the stranded verb, 'seems' suddenly demonstrates that we chose the wrong path of comprehension. Garden-pathing is such a natural comprehension strategy, we are unaware of it until it is interrupted, as it is unintentionally in poor writing, or intentionally in jokes or psycho-linguistic research. The comprehension of texts In addition to the research on sounds, words, and sentences, psycholinguists have also examined the way we process texts. What do we remember of a story that has just been told to us or a letter that we have just read? First of all, with the exception of mnemonists—people who have a rare and uncanny ability to recall texts that they have heard or read—our memory is rather poor for structure but is comparatively very accurate for content. Earlier, we observed that the subjects in the DTC experiments were somewhat hazy about the grammatical form of the sentences they were asked to remember. Passive sentences tend to be remembered as active ones, but usually not the reverse. That is, our syntactic memory may be vague but it is not haphazard; we tend to remember sentences in a form that is actually simpler than the structure which we originally read or heard. If there is not too long a gap in time, subjects can remember that a sentence like (6), to take just one example, was a negative question, but they usually recall it as an active sentence, 'Isn't the dog chasing the cat?' 66 SURVEY The basic content is remembered but not typically the grammar of the sentence. Only when sentences violate our expectations, as in example (8), do we tend to change the meaning into something that more comfortably matches them, like (7). Interestingly enough, even those with supreme memories for texts, the literally one-in-a-million mnemonists, when they finally start to forget the exact wording of a rext, might make minor changes in the words or the grammar, but not in the details of the content. Psycholinguistic research into the comprehension of texts has demonstrated, among other things, that the presence or absence of background information can dramatically affect the way we remember a piece of discourse. In a famous experiment, subjects read a series of paragraphs and, after each reading, they were asked to repeat as much as possible of what they had just read. Read the following example and then try to replicate this study yourself by closing the book and then by attempting to write down as much as possible of what you have just read. With hocked gems financing him, our hero bravely defied all scornful laughter that tried to prevent his scheme. Your eyes deceive you, he had said, an egg not a table correctly typifies this unexplored planet. Now three sturdy sisters sought proof, forging along sometimes through calm vastness, yet more often over turbulent peaks and valleys. Days became weeks as many doubters spread fearful rumors about the edge. At last, from nowhere, welcome winged creatures appeared, signifying momentous success. Not only did you probably experience difficulty in recalling the exact wording and the sequence of sentences in this seemingly incoherent account, you may also have wondered what is was all about. Now give this paragraph to a friend to read and to recall, but before you do so, point out that this is the story of 'Christopher Columbus discovering America'. In the psycholinguistic experiment that contrasted subjects' ability to recall paragraphs like this, those who were given an appropriate title first demonstrated much more accurate recall than those who were not. This suggests that top-down information, which provides general background knowledge about a text, is useful in the comprehension of larger units of language because it helps activate rnupncwrMsinM r mental associations which then assist in overall comprehension and recall. Comprehension concluded Once again we have discovered that an everyday activity that seems to be simple and straightforward is, upon more intensive scrutiny, complex and variable. In the comprehension of speech sounds, we see further evidence that some parts of human language are innate, and do not have to be learned. The perception of major linguistic differences in sounds, such as VOT, is hard-wired into the human brain, and even young children demonstrate the ability to classify very small differences in VOT into one or another phonetic category. This innate ability is extremely useful for children as they grow up hearing their mother tongue, because it allows them to pick up the few significant differences in that particular language and, at the same time, to ignore the many which are insignificant. The research into the comprehension of words has shown that we are very much affected by context, and that our understanding is both facilitated and complicated by the different pieces of knowledge we possess for each logogen. It is clear from the TOT phenomenon that we have access to a dictionary-like memory for words. We can 'search' for a partially-remembered word by comparing and contrasting ocher words which share similar specifications. But our knowledge of and about words is much more extensive: the meaning of a word immediately triggers a spreading activation of associations which help us understand it in many different contexts, and may bring other related words to mind. The grammatical structure of a sentence might initially influence the garden path we choose in trying to understand it, but the greatest influence on sentence comprehension is meaning. We can see this in the experiments with ambiguous sentences because it is clear that ambiguity slows down processing time, but we also observe it in recall. People remember the 'what' that is spoken or written better than the 'how'. Finally, comprehension of larger units of language also indicates the importance of meaning. Texts that fit into a context which we understand and expect are com- SURVEY prehended more quickly and remembered more readily than ones which are presented to us without a context. It is plain that only a complex model of comprehension like PDP can begin to account for the way readers and listeners comprehend the millions of linguistic messages they receive each day. Psycholinguists have to develop a model of comprehension that successfully integrates all the diverse, yet parallel and simultaneous processes that we have examined in this chapter, and obviously such a model will be exceedingly elaborate. It will have to begin with the innate mechanisms for language that are wired into the human mind. It will have to account for the way in which young children rapidly learn to extricate significant phonemes, words, sentence structures, and phrases from the multitude of sounds and sights that besiege them each day. And ultimately it will have to explain how some sort of executive decision-maker in the mysterious garden of the human mind decides when to continue along one path toward understanding, when to abandon it abruptly for a more fruitful alternative, and how to seek almost always successfully an accurate interpretation of the intended message. (Text comprehension passage on page 67 from D.J. Dooling and R. Lachman. 1971. 'Effects of comprehension on retention of prose.' Journal of Experimental Psychology 88:216-22.)