8 •> Psycholinguistics

combine the words into sentences, could get some ideas across, but would lack a system of sufficient precision to convey all her thoughts.

Another way to get a sense of how sentence meaning depends upon structure is to see how the same string of words in the same linear order can convey two different meanings, depending upon the way they are structured. The sentence The man saw the boy with the binoculars is a structurally ambiguous sentence. The sentence can mean either that the man used the binoculars to see the boy or that the boy had a pair of binoculars. Thus, with the binoculars is associated either with the verb saw or with the noun boy. Figure 1.2 illustrates the abstract structures associated with each of the two meanings. In Figure 1.2A with the binoc­ulars is not connected to the boy, but in Figure 1.2B the two phrases are connected. The structures illustrated in these figures reflect the differ­ence in meaning that distinguishes the two interpretations of the sen­tence, which is whether or not the boy possesses the binoculars. These structures, like the ones that will appear elsewhere in this book, are not constructed exactly the way a linguist would construct them. When a linguist draws a representation of the structure of a sentence, it is a the­oretical object, like a drawing of molecular structure by a biochemist. The structures in this book are not nearly as detailed, however, and are meant only to illustrate the particular aspects of sentence structure that need to be focused on.

Linguistic Competence and Linguistic Performance

A grammar and a lexicon are those components of language that allow sounds and meanings to be paired by the construction of sentences. When people know a language, they know its grammar and its lexicon. Psycholinguists call this knowledge linguistic competence. Linguistic competence is a technical term, and it is different from the usual mean­ing of the word competence. Being competent at something usually means that a person has adequate abilities to perform an action with skill, but that is not what is meant by linguistic competence. Linguistic compe­tence has no evaluative connotation; it simply means knowledge of the language that is represented in a person's brain (or mind) that provides a system for pairing sound and meaning. Linguistic performance, on the other hand, is the use of such knowledge in the actual processing of

Beginning Concepts 9

Figure 1.2. Abstract structures associated with the two meanings of the sentence The man saw the boy with the binoculars."

sentences, by which we mean their production and comprehension. Typically, linguists are concerned with the description of linguistic competence and psycholinguists are concerned with the description of linguistic performance. Beyond basic sentence processing, psycho­linguists are also concerned with the actual use of language. After a sentence is processed, it is stored in memory and combined with other

10 Psycholinguistics

sentences to form conversations and narratives. The description of how language is actually used is called pragmatics, and it is often important to distinguish between the grammatical and pragmatic aspects of a par­ticular linguistic event. For example, return to the structurally ambigu­ous sentence The man saw the boy with the binoculars. The sentence can have two distinct meanings, each of which is described by a different structural representation. These two structures are given by the gram­mar and are the result of the application of different syntactic rules. When these sentences are actually used by a speaker and understood by a hearer, only one of the two meanings will be the one intended. Which meaning is intended will be a purely pragmatic issue, deter­mined by the situation and the meaning the speaker wants to convey. The grammar is completely indifferent to the speaker's intent or to the hearer's comprehension. The grammar simply provides structures that are available for the encoding of meaning in sentences. The actual use of those sentences in conversation is a function of encoding and decod­ing processes and pragmatics.

There are several actual processes that must take place when people use language to exchange ideas. Figure 1.3 illustrates the operations of encoding by the speaker and decoding by the hearer. The speaker begins with an idea or a thought she wants to convey to the hearer. In order to do this, she first must translate her thought into a semantic representa­tion of a sentence in her language. Then she must select the words from her lexicon and use her grammar to construct the structure that will convey the meaning she has selected. The words must then be repre­sented phonologically because they are going to be pronounced. Finally, the phonological representation is sent to the motor areas of the speaker's brain and instructions are sent to the organs that will produce speech. The complex speech signal is the result of the precisely timed and exquisitely organized interaction of hundreds of muscles, includ­ing those of the jaw, lips, tongue, vocal cords, and respiratory system. Speech sounds reach the auditory system of the hearer, and he begins the process of reconstruction that is necessary to decode the speaker's message. First, he must reconstruct the phonological representation and discover the speaker's words and their meanings. Then, using the grammatical knowledge that he shares with the speaker, he must recon­struct their structural organization. He then has sufficient information to recover the basic meaning of the speaker's sentence (its semantic rep­resentation) and, finally, her idea or thought. This experience of exchang-

Beginning Concepts 1 1

Encoding Decoding(Speaker) (Hearer)

Idea Idea

'

j

Semantic SemanticRepresentation Representation

'

i

Words + Structure Words + Structure

'

i

Phonological PhonologicalRepresentation Represe*ntation

'Articulatory Speech Auditory

System Signal System

Figure 1.3. Encoding by the speaker and decoding by the hearer.

ing ideas using speech is so commonplace that people never think about the complex cognitive processes that underlie that experience. Like the complex processes underlying most of the activities of living— walking, breathing, sleeping— the activities involved in encoding and decoding sentences are completely unconscious. It is not possible to introspect and experience a piece of the process, like the retrieval of words from the lexicon or the use of one's grammar to create a struc­tural representation of a sentence. Psycholinguists have, however, dis­covered a great deal about these unconscious processes, which are quite remarkable in their speed and complexity.

Notice that in the encoding process an abstract object, an idea, is translated into a physical object, a speech signal. When we say that an idea, like sentence structure, is abstract, we mean that it does not have

12 j*= Psycholinguistics

an observable physical reality Certainly, an idea must have a physical representation deep in the neurological connections of the brain, but it is not observable to the hearer nor is it measurable with any known type of instrument. Speech, on the other hand, is concrete; it is part of observable physical reality. Not only does it have an effect on the hearer's auditory system, it can be recorded and its physical character­istics measured. When the hearer decodes the physical signal, he recov­ers the same abstract object that was encoded by the speaker. Neither the abstract idea nor the physical speech signal is represented by lin­guistic competence. They are not part of the linguistic system, but it is the system that intervenes between the idea and the speech and allows them to be related. The system represents the words and their pronun­ciation and creates the structures that organize them into sentences. These representations join the idea and the speech.

The Speech Signal and Linguistic Perception

The fact that the speech signal is the only physical link between speaker and hearer is a critical psycholinguistic point. The speech signal con­tains all the information required for the hearer to reconstruct abstract structures and ideas, but that reconstruction is essential to the decoding process. To fully appreciate the complexity of this task, it is necessary to understand the relationship between speech and the linguistic repre­sentations that underlie it. In fact, the phonological representation of the speech sounds of the sentence, which is provided by the grammar, is not directly represented in the speech signal produced by the speaker. The speech signal is highly encoded. The phonological representation can be thought of as an idealization of the physical speech sounds. In it individual phonological units are represented, as are individual lexi­cal items. The actual physical speech is very different, however. The phonological units overlap in actual articulation, and the words run together. The speaker may speak rapidly, with chewing gum in her mouth, while a radio is playing. The relationship between the physical signal that the hearer receives and the neat, idealized phonological rep­resentation he must reconstruct is not direct. It is the product of a com­plex ensemble of mental processes. The job of these mental processes is to make use of information from the physical signal to recover the

Beginning Concepts 13

speaker's meaning. They use information about the shared grammar and lexicon of the speaker and hearer to reconstruct the linguistic rep­resentation that joins sound and meaning. Researchers think that those mental processes are executed by neurophysiological operations that are specialized for the perception of speech as a linguistic object.

In every modality people make the distinction between the actual stimulus that impinges on our eyes or ears and the percept that our brain constructs when we interpret that stimulus. A stimulus is never con­sciously available to us; what we are aware of is the percept that the stimulus gives rise to. An example of this process can be illustrated by viewing optical illusions, like the Meuller-Lyer illusion (see Fig­ure 1.4). When examining this illusion, the stimulus that actually falls on our retinas is that of two lines of equal length, but we perceive one line to be longer than the other. The percept of relative length depends not just on the actual length of the lines, but also on the context in which they occur. The fact that these lines are adjoined by angles pointing in different directions affects our perceptual interpretation of their length. The perception of a linguistic representation based on the stimulus of a speech signal is a psycholinguistic activity because it could not be con­structed without the hearer's linguistic competence. Knowledge of lan­guage is necessary for a person to reconstruct, and therefore perceive, a phonological representation of the speech signal. Without linguistic knowledge, a person would be like the dog in the Far Side cartoon (see

Figure 1.4. Mik'iii'i lyir iIIiimoii

52 ^ Psycholinguistics

universal) communication system of human beings is naturally acquired and leads to the next three of Lenneberg's criteria for the biological basis of language.

Language Need Not Be Taught, Nor Can It Be Suppressed

This section combines two of Lenneberg's criteria, which are closely related to one another. Language acquisition in the child is a naturally unfolding process, much like other biologically based behaviors such as walking. Every normal human who experiences language in infancy acquires a linguistic system. Failure to do so is evidence for some sort of pathology. Contrary to the belief of many doting parents, they do not teach their children language. The fact that children need to hear lan­guage in order to acquire it must not be confused with the claim that they need specific instruction of any sort. It is probably the case, how­ever, that children need to experience social, interactive language in order to acquire language. A case study involving two brothers, Glen and Jim, who were the hearing children of deaf parents, illustrates both of these points (Sachs, Bard, & Johnson, 1981). The boys were well cared for and did not suffer emotional deprivation, but they had little experi­ence with spoken language other than from watching television. When discovered by Connecticut authorities, Jim, who was 18 months old, did not speak, and Glen, who was 3 years, 9 months old, knew and used words, but his morphology and sentence structure were virtually nonexistent. Glen would produce sentences such as, "That enough two wing," "Off my mittens," and "This not take off plane." Speech pathol­ogists from the University of Connecticut visited the home regularly and had conversations with the children. They did not attempt to teach them any particular language patterns, but they played with them and interacted linguistically with them. In 6 months, Glen's language was age appropriate and Jim acquired normal language. When last tested, Glen was a very talkative school-age child who was in the top reading group of his class. The story of Jim and Glen illustrates the importance of interactive input for children during the years they are acquiring lan­guage. It also illustrates the fact that specific teaching is not necessary. In the next chapter, the role of caretakers in the acquisition process will be explored further.

The Biological Basis of Language 53

The fact that language learning cannot be suppressed is yet another manifestation of the naturalness of the language acquisition process. If language were not a natural process, if it were more bound to the particular types of linguistic experience a child has, there would be much greater variation in the speed and quality of language learning Ilian is actually observed. In fact, people acquire language at about the same speed during about the same age span, no matter what kind of cultural and social situation they grow up in. Children from impover­ished circumstances with indifferent parental care eventually acquire a fully rich human language, just as do pampered children of affluent, achievement-oriented parents.

Children Everywhere Acquire Language on a Similar Developmental Schedule

There is a remarkable commonality to the milestones of language acqui­sition, no matter where in the world children acquire language. Dan Slobin of the University of California at Berkely has devoted his entire career to the cross-linguistic study of language acquisition and wrote an essay entitled "Children and Language: They Learn the Same all Around the World" (Slobin, 1972). Like the milestones of motor devel­opment (infants roll over, sit up, crawl, and walk at similar ages every­where), the milestones of language acquisition are also very similar. Habies coo in the first half of their first year and begin to babble in the second half. The first word comes in the first half of the second year for just about everyone. In all societies, babies go through a one-word stage, followed by a period of early sentences of increasing length; f inally, complex sentences begin. By the age of 5 the basic structures ofI lu- language are in place, although fine tuning goes on until late child­hood. Children all over the world are sensitive to the same kind of language properties, such as word order and inflection. They make remarkably few errors, but their errors are of a similar type. All chil­dren, for example, overregularize the rule-governed patterns of their language, indicating that the acquisition system of children is special­ized to analyze language in this way. While there is much individual variation in the age at which children acquire aspects of language, that variation is conditioned by individual characteristics of the child rather than by the language he is acquiring or the culture in which he is

54 Psycholinguistics

acquiring it. One would never expect to hear, for instance, that Spanish­speaking children do not use their first word until they are 3, and acqui­sition of Spanish is not completed until puberty. Nor would you expect to hear that infants in Zimbabwe typically begin speaking at the age of 6 months and are using complex sentences by their first birthday. There is clearly a developmental sequence to language acquisition that is independent of the language being acquired, although some features of language are acquired more easily than others. In fact, those aspects of language that are easier and those that are more difficult are similar for all children. All children learn regular patterns better than irregular ones. English-learning children will regularize irregular past tenses and plurals, producing things like "eated" and "sheeps." All children make similar kinds of "errors," no matter what language they are acquiring. Not only is the sequence of development similar for all children, the process of acquisition is similar as well. This is exactly what one would expect if the acquisition of a mental system is being developed accord­ing to a genetically organized, species-specific and species-universal program.

Finally, for children everywhere there seems to be a critical period in the acquisition of their first language. Although the details of this con­cept are controversial, most researchers agree that the optimal period for first language acquisition is before puberty, after which a fully com­plex linguistic system will not develop. The evidence for this comes from reports of so-called "wild children," and more recently from the case of Genie, a California girl who was locked inacloset by an abusive father for the first 13 years of her life (Curtiss, 1977, 1988; Curtiss, Fromkin, Krashen, Rigler, & Rigler, 1974). During that time she was deprived of any linguistic input of any kind. After she was rescued, lin-

^ guists and psycholinguists at the University of California at Los Ange­les worked for years with her to help her acquire English, but to no avail. She acquired words and the ability to communicate verbally, but she never acquired the full morphological and syntactic system of English. These examples of her utterances illustrate the level of her lan­guage ability: "Genie full stomach," "Applesauce buy store," "Want Curtiss play piano," and "Genie have mama have baby grow up." Her hearing was normal, as was her articulation ability. There is some ques­tion as to whether her intelligence was completely normal, but even if it was not, she was not so retarded that her condition would account for her inability to acquire language. Clearly, she had been terribly trau-

The Biological Basis of Language 55

malized for many years, and her emotional development could not Ii.i vi* been normal; however, she was not sufficiently psychologically Impaired to account for her lack of language. Actually, she was quite Iriendly and used language well socially. Her problems were solely in morphology and syntax, the formal aspects of language structure. This tragic story, then, illustrates the claim that after a certain critical period11 ic brain can no longer incorporate the formal properties of language. There are similar stories of deaf people who did not acquire a signed language during the critical period and never developed a fully com­plex signed language. The existence of a critical period for language learning is an argument for the biological basis of language, because iinly biological systems demonstrate critical periods for their develop­ment. Reading and writing, for instance, which are abilities that are cul­lin', illy rather than biologically based, can be acquired at any time in line's life.

While the existence of a critical period for first language learning is lairly well accepted, its relationship to second-language learning is miK’h more controversial. It does seem to be the case that people who learn a second language after puberty almost invariably speak it with (in accent, while those who acquire a second language in childhood do mil.’ The extent to which an adult learner acquires a second language by processes similar to that of a child acquiring a first language is a hotly debated issue in second-language acquisition research. The related l»Hiie of whether the second language is represented in the brain in a »Imilar way as the first language is also a matter of some debate. These tire interesting and important questions; the act of posing them assumes llie biological basis of both first and second language. This brings us toI .enneberg's final point and the issue of the representation of language In the brain.

Language Has Anatomical and Physiological Correlates

The most fundamental biological fact about language is that it is stored In Ilu* brain, and, more importantly, that language function is localized Ina particular area of the brain. This is hardly a new idea, g®ing back rtl least to Gall, the 18th century neuroanatomist who developed the lU'ld of phrenology. He believed that various abilities, such as wisdom,

V> . 'k I ’syi holiiiquislK s

musical ability, morality, and language, were located in different areas of the brain and could be discovered by feeling bumps on a person's skull. Gall was, of course, wrong about the bumps, but it seems to be true that some neurally based abilities, such as language, have specific locations in the brain. The first conclusive demonstration that language was localized in the brain took place in the late 19th century when a French neurologist named Paul Broca presented the first case of apha­sia to the Paris Anthropological Society in 1861 (Dingwall, 1993). Broca had a patient who had received a blow to the head with the result that he could not speak beyond uttering "Tan, Tan," and, thus, Broca called him Tan-Tan. Upon autopsy, he was found to have a lesion in the frontal lobe of the left hemisphere of his brain. Ten years later a German neu­rologist named Carl Wernicke reported a different kind of aphasia, one characterized by fluent, incomprehensible speech (Dingwall, 1993). His patient was found to also have a left hemisphere lesion, farther back in the temporal lobe.

These two kinds of aphasia are still called by the names of the men who first described them, as are the areas of the brain associated with each. Broca's aphasia, also known as a nonfluent aphasia, is characterized by halting, effortful speech caused by damage involving Broca's area in the frontal lobe of the left hemisphere. Wernicke's aphasia, also called a fluent aphasia, is characterized by fluent meaningless strings caused by damage involving Wernicke's area in the temporal lobe of the left hemisphere. These two kinds of aphasias, among others, differ markedly in terms of the grammatical organization of the patient's speech. The speech associated with Broca's aphasia has been character­ized as agrammatic; it consists of primarily content words, lacking syn­tactic and morphological structure. The speech of people with Wer­nicke's aphasia, on the other hand, has stretches of grammatically organized clauses and phrases, but it appears incoherent and meaning­less. In conversation, it appears that people with Broca's aphasia com­prehend what is said to them, while people with Wernicke's aphasia do not. Thus, a general clinical characterization has been that people with Broca's aphasia have more of a problem with speech than with auditory comprehension, whereas people with Wernicke's aphasia produce flu­ent, well-articulated but meaningless speech, and have problems with auditory comprehension. Figure 3.1 is a sketch of the left hemisphere of the cortex of the brain, with Broca's and Wernicke's areas indicated. It is interesting to note that Broca's area is located near the motor area of the

The Biological Basis of Language J*= 57

Figure 3.1. The left hemisphere of the human cerebral cortex (side view). BrocaS and W< micke's areas are the primary language areas. The former is located in the Frontal lobe; il ii' latter in the Temporal lobe, which also contains the Auditory area. They are divided by il ii ■ Sylvian fissure. The Motor and Sensory (pain, touch, temperature, etc.) areas are in the I n mtal and Parietal lobes, respectively, and are divided by the Central sulcus. The fourth Ii >l )(■ is the Occipital, which contains the Visual area.

cortex, while Wernicke's is near the auditory area. Although this is true, the difficulty Broca's patients experience with speech is not because they have motor problems; it is purely a language problem. Similarly, the auditory comprehension difficulties associated with Wernicke's aphasia are not an auditory problem. The hearing of people with this aphasia typically is fine, but they have a language problem. It is, of course, also possible for motor problems to accompany Broca's aphasia and for hearing problems to accompany Wernicke's aphasia. The point is that these are not the primary causes of the difficulties observed in those two kinds of aphasias.

Deaf people who use signed languages can also become aphasic with damage to the left hemisphere. Their signs are nonfluent, halting, and agrammatic. This is true, despite the fact that they have no motor

disability in their hands and can use thorn in ovoryd.iy tasks with no difficulty (Poizner, Klima, & Bellugi, 1987). Tho fac t that signors become aphasic is dramatic confirmation of the fact that signed languages not only have all the formal properties of spoken language, but are simi­larly represented in the brain. It also demonstrates that the neurological damage which produces aphasia impairs language systems, rather than motor systems.

Nothing is ever quite as simple as it might seem initially. Psycho­linguists studying the comprehension ability of people with Broca's aphasia discovered something very interesting. People with Broca's apha­sia had no difficulty in understanding sentences with relative clauses, such as The apple the boy is eating is red, but had difficulty with The girl the boy is chasing is tall (Caramazza & Zurif, 1976). Both sentences are constructed of common words and identical structures; there is, how­ever, a profound difference between them. The second sentence cannot be understood unless the hearer analyzes the relative clause structure; otherwise, one cannot know who is chasing whom and who is tall. The first sentence is very different, however. One need not process the rela­tive clause, because boys eat apples, not the other way around; more­over apples, not boys, are red. Real world knowledge allows a person to successfully guess the meaning of the first sentence, but not the sec­ond. Comprehension of the second requires an intact syntactic process­ing system. This result suggests an explanation as to why people with Broca's aphasia seem to have little trouble with auditory comprehen­sion in conversational contexts. In ordinary conversation with people one knows well and with whom one shares a great deal of real world knowledge, one can understand much of what is said without having to do a full analysis of sentence structure. The reasonably successful comprehension of people with aphasia seemed to result from an ability to figure out the meanings of sentences in discourse and compensate for an impaired grammatical processing system. The question remains, of course, as to whether their grammatical problems are a result of an impaired linguistic competence or are the result of difficulty in using that competence to produce and understand speech. It is very difficult to answer this question experimentally, but some researchers have found people with agrammatic aphasia whose metalinguistic skills with respect to syntax are better than their ability to perform syntactic processing (Linebarger, Schwartz, & Saffran, 1983). This would suggest that the performance system is more impaired than the underlying grammar.

I M l 1 n i M l l H J H ,11 I M N n H I l i l M l j U t H j r i * > f

Nciuvliiiguislics is llio study ol tho representation of language in the Inain, and neurolinguisls are currently very interested in such issues.

Aphasia is not a simple or clear-cut disorder. There are many dif­ferent kinds of aphasia in addition to those classified as fluent and non-II nont, and many different behaviors that characterize the various clin­ical types of aphasia. Furthermore, much more of the left hemisphere is involved with language than just Broca's and Wernicke's areas. The area all along the Sylvian fissure, deep into the cortex, is associated with language function. Moreover, people with aphasia differ greatly in the severity of their symptoms, ranging from mild impairment to a global aphasia where all four language modalities— auditory and reading comprehension, and oral and written expression— are severely impaired.

The Lateralization of Language

lb say that language is lateralized means that the language function is located in one of the two hemispheres of the cerebral cortex. For the vast majority of people, language is lateralized and it is located in the left hemisphere. However, in some people language is lateralized in the right hemisphere, and in a small percentage of people language is not lateralized at all, but seems to be represented in both hemispheres. The hemisphere of localization is related to handedness, left-handed people being more likely to have language lateralized in the right hemi­sphere. Exactly why this should be the case is unclear, but it is true that the right side of the body is controlled by the left motor and sensory areas, while the left side of the body is controlled by the right motor and sensory areas. Thus, left-handed people have right-dominant motor areas, while right-handed people have left-dominant motor areas. A large study by Rasmussen and Milner (1977) gives a sense of the rela­tionship between handedness and lateralization in the population. In this study, 262 people were given the Wada test, which can determine in which hemisphere of the brain a person's language is localized. This is a very important concern if a person is about to undergo brain sur­gery; the surgeon must know where the person's language function is localized so the surgery will not leave him or her aphasic. In the Wada test, sodium amytol is injected in such a way that it will enter one of the two hemispheres of the brain. The patient is asked to Aunt or name

I I » . ■» rsyt ГКШМ(|||1М1( s

pictures presented on an overhead screen. Hoc.hiho each hemisphere controls the functioning of the opposite (conlmlalcrnl) side of the body, the injection produces paralysis on the side of the body opposite from the affected hemisphere. The injection also disrupts verbal behavior, only briefly if the nondominant hemisphere has been injected, but for several minutes if it has been the dominant hemisphere. Of the right- handed people in the Rasmussen and Milner study, 96% of them had language lateralized in the left hemisphere, and only 4% in the right. Of the left-handed people, however, only 70% were left lateralized, 15% were right lateralized, and 15% had language function located in both hemispheres. It is evident that the majority of left handers are left later­alized, but there is a slightly higher probability that they will have lan­guage located in either the right hemisphere or in both.

There have been many other demonstrations of the specialization of the left hemisphere for language. One such example is a procedure called brain mapping, originally developed by Penfield and Roberts in the 1950s (Penfield & Roberts, 1959) and described extensively by Oje- mann at the University of Washington (Ojemann, 1983). This procedure is also carried out prior to neurosurgery to provide a detailed descrip­tion of exactly which areas of the brain should be spared by the sur­geon's knife in order to prevent an aphasic outcome. In this procedure, patients are given a spinal anaesthetic so they will be able to communi­cate with the clinician. The skull is opened and the brain is exposed; because the brain itself has no nerve endings, this is not a painful pro­cedure. Various areas are marked along the surface of the brain, and a brief electric current is administered at the same time the patient is per­forming a verbal task. If that area is associated with language function, the electric current will produce a temporary aphasic-like episode. In the procedure, the patient is shown a picture of a ball and instructed to say, "This is a ball." At the moment the word ball is about to be pro­duced, a mild electric current is applied to a small area of the left hemi­sphere. If that is a language area, the patient will not be able to say "ball." If it is not in the language area, there will be no interruption in speech. Surgeons do not cut within 2 centimeters of the areas identified in this manner. In the patients he studied, Ojemann found that they had language areas located in Broca's area in the frontal lobe, in Wernicke's area in the temporal lobe, and all around the Sylvian fissure in the left hemisphere, but nowhere in the right hemisphere. Further, there seemed to be some areas that were specialized for word naming and

l l i r limb h |I( ,il M.iMs пГ I ,iM (|ii.ic|t’ . * M

(tillers that were specialized lor syntax (although most areas included holh abilities). Seven tlreek-linglish bilinguals participated in Oje- nmini's study; for each of them there were a few areas in which Greek, hill not Hnglish, was located, and others where English, but not Greek, w.is located (although many areas overlapped). This finding addressesI he i m portant question of how two or more languages are located in the hr.)in. It is well known among speech-language pathologists that there иге reports of bilingual people with brain damage becoming aphasic in one but not both of their languages.

Л particularly fascinating demonstration of the lateralization of l.mguage function comes from patients who have had a surgical proce­dure called commissurotomy, in which the two hemispheres of the cortex .in’ separated by cutting the corpus callosum, a thick sheaf of nerve fibres that join the two hemispheres. This procedure is performed in cases of severe epilepsy in order to prevent the electrical impulses that cause seizures from surging from one hemisphere to the other. Roger Sperry( 1968) received the Nobel prize for work with people who have had this surgery, and many other people have worked with them as well (Gaz- zaniga, 1970). To understand what they have found, it is necessary to recall that the right side of the body is controlled by the left side of the brain, and vice versa. The neural pathways that control the motor and sensory activities of the body are below the area severed by the com­missurotomy, so the right motor areas still control the left hand and the right sensory areas receive information from the left side of the body. The right hemisphere of the brain cannot, however, transfer informa­tion to the left hemisphere, nor can it receive information from the left hemisphere. Suppose that a person has language lateralized in his left hemisphere and has had such a surgical procedure. If his eyes are closed and a ball is placed in his left hand, he will not be able to say what it is. However, he would be able to select from an array of objects the object that he had held in his hand. The right hemisphere has knowledge of the identity of the ball, but it lacks the ability to speak, so the person is unable to name it. If the ball is placed in his right hand, he is able to name it, just as any person would be able to do with either hand. If a person were to close her eyes and have someone put a ball in her left hand, the information that it is a ball would register in her right hemisphere, then her right hemisphere would send the information to the left hemisphere, which would name it. The step that is missing for the person who has a "split brain" is the information tranffer from the right to the left hemisphere.

It is also possible to present information to eitlu-r (he left visual field, which sends information to the right hemisphere, or to the right visual field, which sends information to the left hemisphere. The left visual field is not the same thing as the left eye; it is a bit more compli­cated than that. Information in the left visual field comes from both eyes (as does information in the right visual field), but what is of inter­est here is that the information from the left visual field goes only to the right hemisphere, and information from the right visual field goes only to the left hemisphere. If a person with a split brain is presented with a picture of a spoon in the left visual field, he will not be able to name it, but he will be able to select a spoon from an array of objects with his left hand. This shows that the right hemisphere recognized the spoon, although it could not name it. The fact that visual information can be presented to one or the other hemisphere has allowed psycholinguists to study in some detail the kinds of linguistic tasks each of the hemi­spheres can perform. While the right hemisphere is mute, it can recog­nize simple words, suggesting that there is some sort of lexical rep­resentation in the right hemisphere. However, there seems to be no representation of formal aspects of language. The right hemisphere can­not rhyme, suggesting that it does not have access to the internal phonological structure of lexical items. Neither does the right hemi­sphere have access to even simple syntax. In an experiment that tested whether subjects could match simple sentences with pictures they had seen when presented to the right hemisphere, the subjects could not dis­tinguish between The boy kisses the girl versus The girl kisses the boy, The girl is drinking versus The girl will drink, and The dog jumps over the fence versus The dogs jump over the fence (Gazzaniga & Hillyard, 1971). Thus, while the right hemisphere may possess some rudimentary lexical infor­mation, it is mute and does not represent the phonological, morpholog­ical, and syntactic form of language.

Further evidence of the dominance of the left hemisphere for lan­guage comes from studies of dichotic listening. In this kind of experi­ment, subjects are presented auditory stimuli over headphones, with different inputs to the two ears. For instance, the syllable ba might be played into the right ear, while at the same exact time da is played to the left ear. The subject's task is to report what was heard. On average, stimuli presented to the right ear are reported with greater accuracy than the stimuli presented to the left ear. This is known as the right-ear advantage for language. It occurs because a linguistic signal presented to

The Biological Basis of L.incju.iqe 63

llu> right ear arrives in the left hemisphere for decoding by a more11 i a r t route than does a signal presented to the left ear. From the left ear, the signal must travel first to the right hemisphere, then across the cor­pus callosum to the left hemisphere. (Kimura, 1961, 1973). Thus, infor- n i.i I ion presented to the right ear is decoded by the left hemisphere «■.irlier than the information presented to the left ear. The right-ear advantage exists only for linguistic stimuli. Nonspeech signals produce no ear advantage, and musical stimuli demonstrate a left-ear advan­tage (Kimura, 1964).

Lateralization apparently begins quite early in life. Evidence sug­gests that the left hemisphere is larger than the right before birth, and infants are better able to distinguish speech from nonspeech when the stimuli are presented to the left hemisphere (Entus, 1975; Molfese, N73). Early language, however, appears not to be lateralized until the age of about 2. If the left hemisphere is damaged in infancy, the right hemisphere can take over its function. This ability of parts of the young brain to assume functions usually associated with other areas is called plasticity. An infant or young child who suffers left hemisphere damage is far more likely to recover without suffering aphasia than an adult, whose brain is far less plastic. Even children who have undergone sur­gery in which the left hemisphere is removed can develop quite good language functions. However, studies have shown that such children are deficient in the formal aspects of language— morphology and syn­tax. Thus, the right hemisphere may be limited in its plasticity in that it i an not incorporate the structural analytical aspects of language associ­ated with the left hemisphere (Dennis & Whitaker, 1976).

Evidence of Language Processing in the Brain

I ’sycholinguists who study how the brain processes language are very interested in the ability to record event-related potentials (ERPs) during language processing. ERPs are changes in the electrical patterns of the brain that can be timed and associated with the processing of various kinds of linguistic stimuli. In typical ERP experiments, sentences are presented to subjects word-by-word so it will be easy to ifvatch ERPs with the areas of the sentence to which they are responding (Osterhout, 1994). Two kinds of HKI’s are of particular interest. One is known as the

!!C H A P T E R 5 ' C I

The Speaker: Producing Speech

The processes that underlie the production and comprehension of speech are information-processing activities. The speak­er's job is to encode an idea into a spoken sentence that will

carry all the information the hearer needs to decode the speech signal ,ind recover the intended message. These activities are essentially mir­ror images of one another. The speaker knows what she intends to say; her task is to formulate the message into a set of words with a structural organization appropriate to convey that meaning/ then to transform the structured message into intelligible speech. The hearer, on the other hand, experiences only the speech produced by the speaker and must reconstruct the intended meaning. This chapter and the next dis­cuss the information-processing operations performed—rapidly and unconsciously—by the speaker and the hearer, as well as the mental representations constructed by those operations.

In the last 15 years, according to Bock (1991), work in the area of language production has made the transition from a psycholinguistic hobby to a central concern in the study of language performance. The production of a sentence begins with the speaker's intention to com­municate an idea or some item of information, referred to by Levelt (1989) as a preverbal message, because at this point the idea has not yet been cast into a linguistic form. This is the task of theformulttor, making

91

reference to the lexicon and Ihe grammar ol the language shared by tht speaker and hearer. Finally, the information must be carried in a speech signal that is produced fluently, at exactly the correct rate, with pre­cisely the correct prosody. There are a number of steps to this process, each associated with a distinct representation of different forms of tha sentence and each carrying its own particular type of information. Fig­ure 5.1 presents a summary of the processing operations performed by the speaker.

Idea

Sentence Meaning

ILexical Retrieval

andCreation of Structure

IApplication o f Phonological

andMorphological Rules

IPhonological Representation

.......I.......Instructions to Articulators

IProduction of Speech

Figure 5 .1 . Summary of the processing operations performed by the speaker

I hr '.| irviM I'mi Iik IIK| h . » V I

The first step is to create a linguistic representation of a sentence meaning that will convey the speaker's thought. Since the meaning of a »rnlence is a function of the words in it and their structural organiza- ti< in, the next encoding stage creates that ensemble of words and struc- Iiiiv. I he appropriate lexical items must be retrieved from the speaker's Internal lexicon, and the appropriate structure must be generated by wyntactic and morphological rules. This will produce a set of words m| »i*lled out in phonemes (as they were stored in the lexicon), hierarchi­cally organized, with inflectional morphemes specified, such as [past], |plural], [third person singular], and so on. Phonological and morpho- phonological rules then apply to produce a final string of phonological elements that will specify the exact pronunciation of the sentence, including its prosodic characteristics. This representation is translated into instructions to the vocal apparatus from the motor control areas of the brain, and then neural signals are sent out to the muscles of the lips, tongue, larynx, mandible, and respirators to produce the actual speech signal.

Recall that in Chapter 1 language and speech were distinguished, .is were language and general cognition (intelligence). This was done in order to show that language is a distinct, autonomous system that inter­acts with other systems of the language user, but which can be described without reference to them. The interaction of linguistic and nonlinguis­tic systems is a continuing theme of this book, because it is the key to understanding psycholinguistic processes. Psycholinguists disagree on many of the details regarding the nature and the degree of the inter­action between linguistic and nonlinguistic systems, but the fact that they do interact is relatively uncontroversial. One can see an excellent example of this phenomenon in sentence planning. An idea is a product of the speaker's cognition and intellect. Speech is a complex motor activity that is the product of vocal tract and respiratory physiology. Language enables the speech to transmit the idea. All of the representa­tions inside the dotted lines of Figure 5.1 are products of the speaker's language. The theory of language provides a vocabulary and a frame­work to represent words, structure, morphemes, and phonological seg­ments. The elements outside the dotted lines are not part of linguistic theory and have no linguistic representation. Thus, the speech produc­tion mechanism is a sequence of processes that links two nonlinguistic and quite distinct representations— an idea and a speech signal— using the speaker's knowledge of language.

Planning Speech Before It Is Produced

The construction of a sentence involves a series of distinct operations and representations. The final representation consists of all the phonetic detail that will direct the actual production of the sentence, exactly like a phonetic representation. In this representation the phonological seg­ments are arranged in a linear sequence, one after the other, as if they were waiting in the wings of a theater preparing to enter the stage. How do researchers know that the phonemes of a sentence are repre­sented before they are produced? One way psycholinguists have inves­tigated this question is by looking at the speech errors people make. The following examples and analyses of speech errors and levels of production planning are based on the work of Garrett (1980a, 1980b, 1988) and Fromkin (1971, 1980, 1988). Speech errors are often called "Spoonerisms," named after a 19th century professor at Oxford in Eng­land who was reputed to make hilarious speech blunders, such as say­ing "Queer old dean" when he intended "Dear old queen," or saying "You have tasted the whole worm" for "You have wasted the whole term." Psycholinguists have learned a great deal by analyzing occur­ring speech errors. Speech errors can be classified as the following:

Perseveration (e.g., "I can't cook worth a cam" for "I can't cook worth a damn")

Anticipatory (e.g., "taddle tennis" for "paddle tennis")

Exchange (e.g., "hass or grash" for "hash or grass")

In the perseveration error, a speech sound is substituted for an earlier one, while in the anticipatory error a speech sound that has not yet been ISt produced intrudes. The exchange error is a combination of the two in that two speech sounds actually change places.

The occurrence of errors of this type demonstrates that there is a level of representation where the phonological elements are repre­sented segmentally. Such errors can tell us much more about the psycho­logical reality of linguistic representations before sound is produced. Speech errors never create phonemes that are not part of the phonemic inventory of the speaker's language, nor do they create impossible words— that is, words that violate the phonotactic rules of the speaker's language. Neither do speech errors violate the phonological rules of the

Mpc.iker's language. Victoria I Tom kin (1973) wrote an article about Ihis research in the Scientific American magazine, titled "Slips of the 'longue." In it she pointed out that a slip of the tongue while pronounc­ing the title might result in "Stips of the Lung," but one would never expect to hear an error like "Tlips of the Sung" because "tlips" violates the phonotactic constraints of English. There are many other linguisti­cally based regularities connected with such speech errors. Conso­nants and vowels never substitute for one another; substitutions and exchanges take place only between elements that are linguistically similar.

Probably the most interesting aspect of these speech errors is that they show (in the case of exchange and anticipatory errors) that there is .i mental representation of phonological segments some time before they are produced. For "reading list" to come out "leading list," there must be a representation of the /1/ in "list" when the /r/ in "reading" is supposed to be pronounced. The majority of errors of this type occur within clauses, suggesting that speech is organized in clause-sized bundles before it is produced.

Some word-exchange errors also provide evidence for the fact that words, like phonological elements, are represented as separate units. Thus, observed errors were "I left the briefcase in my cigar" for "I left the cigar in my briefcase," and "rubber pipe and lead hose" for "rubber hose and lead pipe." Exchanges never occur between content words and function words and are usually limited to words of the same gram­matical class, nouns in the case of the above examples.

Word-exchange errors can also provide information about other aspects of the representation of a sentence prior to production. It is important to remember that the prosodic information about how a sentence is to be pronounced is provided in the final representation. This includes information about which words in the sentence will be stressed. In the clause "When the STORY hits the paper," primary stress is placed on story (indicated by capital letters). A word-exchange error resulted in this sentence being produced as "When the PAPER hits the story." Primary stress was on paper, which ended up in the same posi­tion in the sentence as story should have been. This demonstrates that the stress was applied based on the structure of the clause, rather than based on being associated with a particular lexical item. Furthermore, the stress (being associated with the structure) did not move with the lexical item. If that had happened, the result would have $een "When

Vfi > » |\y< IUllllK|lllS(l( S

the paper hits the STORY." Another example <>l this N.11110 phenomenon is when “Stop beating your HEAD against a brick wall" was intended and was produced as "Stop beating your BRICK against a head wall." Again, the stress retained its position in the sentence, demonstrating that it is a property of the form of the sentence and not of the individual words. Of course, the actual production of the sentence requires that the word be stressed. The representation of the stress, however, is a property of the sentence.

Exchange errors can also involve syntactic units larger than indi­vidual words. Such errors provide information about the structural representation of sentences. Constituents that are larger than words, but which are units in the hierarchical organization of the sentence, can exchange with one another. For example, a speaker intended to say, "There's a small restaurant on the island," but produced instead, "There's an island on the small restaurant." The noun island b“ changed places with small restaurant, which is an entire noun phraaq» Thus, a constituent larger than an individual word has moved. Move­ment of two words which are not part of the same constitutent is never observed. An error such as "There's island restaurant on the a small" would never be produced. In speech errors, syntactically defined con­stituents are moved, and the resulting sentences are always structurally well-formed sentences of English.

Word-exchange errors can also demonstrate the existence of a level of representation where bound morphemes are represented separately from their stems. A person intended to say, "You ended up ordering," and instead said, "You ordered up ending." The words end and order were exchanged without the bound morphemes -ed and -ing moving along with them. The error was not "You ordering up ended," as it would have been if the bound morphemes and the stem had formed a unit at the time of exchange. This demonstrates that there is a level of representation at which the stem and bound morpheme are distinct entities. It is at such a level that this word-exchange error occurred. Moreover, it is an example of the fact that speech errors rarely produce structurally bizarre sentences (although their meaning may be odd).

Similar kinds of errors demonstrate not only that there is a level of representation at which stems are represented separate from bound morphemes, but also that those morphemes are represented according to their morphological category, before the morphophonological rules operate to specify the phonetic form by which the morpheme will be

realized. For example, .someone intended to say, "We cooked a roast," .nid instead produced, "We roasted a cook." (This might initially .if »pear to contradict the observation that only words of the same gram­matical class are exchanged, since cook is a verb and roast is a noun.I lowever, both words can be either a noun or a verb, so the example is not a contradiction.) The past-tense morpheme differs in the way it is pronounced depending upon the final segment of the verb to which it is attached. The past-tense morpheme on cook would be the [t] sound, while on roast it is [Id], In the speech error cited above, the past-tense marker is "spelled out" according to morphophonological rules attach­ing it to roast. Thus, it is clear that roast and cook were exchanged before the morphophonological rules applied. The exchange error resulting in I he sentence "We roasted the cook" thus provides evidence for a level1 >1 representation as shown in Figure 5.2. The words were exchanged at ,1 processing level before the morphological rules had applied to create a phonetic form for the past tense. If the exchange error had occurred

I ihonemic rules.

9H , * I'syi hi ilu k |i ilslIt s

at a later processing stage the sentturwM» would huve come out as "W troast a cooked." Such a speech error would never occur.

Similar structural information is revealed by the following speech error. Someone intended to say, "If you give the infant a nipple," but instead said, "If you give the nipple an infant." In this example, nipple and infant have been exchanged before the morphophonological rule specifying the pronunciation of the indefinite article has applied. The article would have been pronounced "a" before "nipple," but instead became "an," which is detemined by the initial segment of "infant." Had the exchange error occurred after the application of the morpho­phonological rule, the resulting sentence would have been "If you give the nipple a infant."

Evidence from speech errors demonstrates that before a sentence is spoken it i§ represented structurally and morphologically, and then, following application of phonological and morphophonological rules, it is represented as a sequence of phonological segments with prosodic structure specified. At the time the structure is represented, content words are specified, but as the indefinite article example showed, func­tion words may not be phonologically represented if their pronuncia­tion relies on the application of morphophonological rules.

The words are selected by a process called lexical retrieval. Remem­ber that the lexicon is a dictionary of all the words a speaker knows; it carries information about the meaning of the word, its g ram m atie i class, the structures into which it can enter, and the way the word i|g pronounced (its phonological representation). A word can be retrieved using two different kinds of information— meaning or sound. Thfc speaker begins with the meaning to be communicated and has the tasjgp of selecting a word that will be appropriate for the desired message The word must also be of the appropriate grammatical class (noun, verb, etc.) and must be compatible with the structure that is being ca§g5* stracted. It is most certainly not the case that the structure is con­structed before the words are selected, nor are the words selected before the structure is constructed. In fact, the words and the structure are so closely related that the two processes must take place simultane­ously. Ultimately, however, the speaker must retrieve a lexical item thai.^ will convey the correct meaning and fit the intended structure. This means that a speaker must enter the lexicon via information about grammatical class, structure, and meaning, and retrieve the phonologr’ ical form of the required word* The hearer's task, which will be dis-

Ihr ‘,| irMlri 1'k kIi ii lh(| S|)fc( 11 . * 99

uissod in detail in the next ilmpler, i» the mirror image of the speaker's, rhe hearer must procemi information about the sound of the word and enter his lexicon to discover its form class, structural requirements, and meaning. Important psycholinguistic questions concern the organiza­tion of the lexicon and how it is accessed for both production and com­prehension. While psycholinguists know a bit about these operations,I heir speed remains a mystery. In conversational speech, 20 to 30 pho­netic elements are produced each second. Thus, a word consisting of10 phonetic segments would be produced in about half a second. Because this includes time to move the articulators, it is impossible to estimate how much of this time is actually spent in lexical retrieval, but clearly I lie process must be extremely rapid. According to Miller and Gildea ( 1987), adults with a high school education know around 40,000 words. 411 the different versions of a single word count as one word. For example, write, writer, writes, written, and writing together count as one word. If one adds to that total another 40,000 proper names of people and places, the adult lexicon is estimated to contain around 80,000 words.11 each word a person uses must be retrieved from a bank of 80,000 in less than half a second, it is obvious that the processes employed in lex­ical retrieval must be extremely efficient.

One way the lexicon is organized is by frequency of use. More com­mon words are accessed more rapidly by both the speaker and hearer. Studies of pauses and hesitations in speech have shown that hesitations often occur before low-frequency words (Levelt, 1983). Words are also organized by their meaning, so words that are similar in meaning or which are close associates are stored near one another. Again, speech errors can give some insight into this organization. It is extremely com­mon for a word retrieval error to result in the selection of a semantically and structurally similar word. For example, a speaker may say, just feel like whipped cream and mushrooms," when he intended to say "strawberries" instead of "mushrooms." Another example is when a person said, "All I want is something for my elbows," when "shoul­ders" was intended instead of "elbows." In each case, the speaker has erroneously selected a word that is of the same grammatical class (nouns in both examples) and that shares many aspects of meaning with the intended word. This kind of error is very common and is prob­ably responsible for many of the so-called "Freudian slips" that people make, such as "I hate, I mean love, dancing with you." However, rather than representing a repressed desire for mushrooms or a secret loathing

of one's dance partner, these errors art* more likely the result of words that share many features of grammatical form and meaning being orga­nized together. Words that are the opposite of one another actually share a great many aspects of meaning, such as hate and love. These words are both nouns that refer to internalized feelings one person can have about another; the only difference between them is that they refer to distinct * (and opposite) feelings. Speech errors will often involve the production of "forget" when "remember" was intended, "give" instead of "take," and so on.

Similar kinds of errors occur based on phonological structure. Thus, words that sound alike are implicated in speech errors. For example, a speaker said, "If you can find a garlic around the house," when the speaker intended gargle, and another said, "We need a few laughs to break up the mahagony," when the intended word was monotony. In these errors, the grammatical class is the same, but the meaning is com­pletely different from the intended word. This suggests that words are also organized by phonological structure, forming "neighborhoods" of words that sound similar.

An unexplained mystery of lexical retrieval that has interested psy­cholinguists for decades is the tip o f the tongue phenomenon (Aitchison, 1987; Brown & McNeill, 1966). This common phenomenon occurs when people know the word they need, but they cannot quite retrieve it. It is a very uncomfortable mental state, and when people experience it, they say, "I've got that word (or name) right on the tip of my tongue!" The interesting feature of the tip-of-the-tongue state is that people always know something about the word they are unsuccessfully searching for. One can often think of its initial or final sounds or letters, how many syllables it has, where primary stress is located, and even words that sound similar. For example, I recently had the name of an acquaintance on the tip of my tongue, and I could recall only that it was a name that has two different spellings (it was Janice/Janis). While no one really understands the tip-of-the-tongue state, it does demonstrate that when people enter the lexicon through meaning, in order to produce a word, a great deal of information may be available even if the entire represen­tation of the word is not retrieved. This is, of course, a rare occurrence, as are lexical errors. Usually lexical retrieval produces an appropriate set of words required for the speaker's sentence.

The sentence-planning process, then, ends with a sentence repre­sented by structure, prosody, and lexical items to which phonologi-

| f | P l | № R P t ’ | I | I II |l M II M. | . r

i i»l iind morphophonoloKic.il rules have applied, in order to create a detailed phonetic representation of the sentence, which now needs to he transformed into sound.

Producing Speech After It Is Planned: The Source-Filter Theory of Vowel Production

The abstract phonetic representation of the speaker's sentence is sent to I lie central motor areas of the brain, where it is converted into instruc- I ions to the vocal tract to produce the required sounds. The production of speech, though commonplace and unconscious, is an incredibly t omplex motor activity. It involves over 100 muscles moving in precise synchrony to produce speech at the rate of 10 to 15 phonetic units per

f second (Liberman, Cooper, Shankweiler, & Studdert-Kennedy, 1967).I Inspiration during speech is very different than it is during silence,

Invause during speech the air from the lungs must be released with exactly the correct pressure to maintain the fundamental frequency {p i tch) of the voice during stretches of speech. The respirators, working with the muscles of the larynx to control the rate of vibration of the vocal folds, maintain the fundamental frequency of the voice and pro­vide the variations in frequency and stress necessary for the prosody of speech. Muscles of the lips, the tongue, and other articulators must be carefully coordinated, leaving little room for error. Much precision of planning is required. For example, to make the vowel sound "oo," writ­ten phonemically as /u/, the lips must be rounded and the larynx slightly lowered. Different sets of nerves enervate the lips and the lar­ynx. Furthermore, impulses travel at different rates down those two sets of nerves, so one impulse must be sent a fraction of a millisecond sooner than the other. This is an example of the level of planning car­ried out by the central planning system in the brain.

To begin to understand how speech is produced, one can examine the vowels /i/ (pronounced "ee"), /a/ ("ah"), and /u/ ("oo"). Fig­ure 5.3 is a diagram illustrating the speech-producing organs. Speech consists of a sound (generated by the vocal folds) being passed through a filter (the vocal tract). This is called the source/filter theory of speech production. Air comes from the lungs, passes through the larynx, and causes the vocal folds to vibrate. The frequency with which they vibrate is called the fundamental frequency of the voice. This is what makes a

The Hearer: Speech Perception and

Lexical Access

T he previous chapter dealt with the operations performed by the speaker to encode a mental message via language into a physical signal that is accessible to the hearer. The hearer's

hisk is diagrammed in Figure 6.1. The hearer's task is essentially the mirror image of the speaker's task, because the hearer must reverse each of the processes performed by the speaker in order to recover the intended meaning. The first task of the hearer is to decode the physical speech signal in order to reconstruct the phonological representation of the message. Using that knowledge, the hearer enters his lexicon to recover the meanings and structural details of the words in the mes­sage. Next, the hearer reconstructs the structural organization of the words and then has the information necessary to recover the speaker's meaning. A description of these operations is the subject of the next two chapters. Speech perception and lexical access are addressed together because they interact in interesting ways. Both phonetic ele­ments and words must be extracted from a continuous, nonsegmented, highly coarticulated signal. There are no spaces between the phonetic units, and there are no spaces between the words. Thus, some of the same problems exist for the speech perception system as for lexical access.

0

113

Basic Sentence Meaning

Structural Processor

Lexical Processor

Phonological RepresentationA

Speech Perception System

ÏSpeech Signal

Figure 6 .1 . Sentence processing operations performed by the hearer to recover the speakers meaning.

Perceiving Speech

The hearer has the role of the inspector in the Hockett quotation from the previous chapter. The phonetic "eggs" have been mangled and mixed together by articulatory processes; it is the hearer's task to iden­tify from the resulting mess of the speech signal what the original pho­netic elements were. There are three features of the speech signal that must be dealt with by the speech perception system.

First, the speech signal is continuous and must be segmented—fifS!* into phonetic units, then into words. Second, because of coarticulatMJH, the speech signal is characterized by the parallel transmission of in fa ^ mation about phonetic segments. Figure 6.2 shows how the information about all three phonetic units in the word bag is distributed across the word. The vowel /ae/ influences the pronunciation of the entire word.

ІИРІІРЙІРІ S|I«»W II ГПІ rpiltill .111(1 І Г'Лі і іИ лм і i-n . -

I'lir influence of the Inllinl /b/ continues through the vowel and into ll«* In-ginning of the /g/. The influence of the /g/ begins at the offset nl llio /b/ and continues throughout the vowel. Thus, there are many h u m s in the center of the word (between the dotted lines) where the aptvi h signal carries information about all three of its phonetic units. This is an example of parallel transmission. A more accurate name for Ihls would be simultaneous transmission of information. Many slices of the speech signal are simultaneously carrying information about a number <>/ phonetic units. The speech perception system must sort out all that Information and figure out what the units are.

A third feature of the speech signal is its variability, also known as11 if lack o f invariance. While the abstract, phonetic representation of a particular phonological element does not vary, each time it is actually produced, the sound may vary greatly. Many factors contribute to the liid that the same sound, the same syllable, or the same word are never pronounced exactly the same. First, there is variation among speakers. I )ifforent people produce speech with different fundamental frequen­cies and have vocal tracts of slightly different shapes. The exact formant frequencies of vowels and formant transitions of consonants are very different from person to person, but their patterning will be similar for everyone. The perceptual system must rely on the relationship among

в æ G

Figure 6.2. Illustration of parallel transmission of phonetic information. Note. Adapted Imm "The Grammars of Speech and Language," by A. M. Liberman, 1970, Cognitive Psy- ( hology, 1, pp. 301-323. Copyright 1970 by Cognitive Psychology. *

I I (> . * I’vyi III llllKjl HMIl \

acoustic elements, such as the fact that the iirnt iormant oi /a/ is W|. relative to /i/ and /u/, no matter who Is speaking. Second, there il tremendous variability within the speaker. People speak fast and slow, they talk with chewing gum in their mouths; they mumble; they shout/ they talk from the next room and while music is playing. All of the:* things affect the quality of the speech signal and make a single word sound very different each time it is uttered, even by the same speakef. There is also contextual variation in the pronunciation of phonetic uni ̂and individual words. Just as coarticulation creates context effects phonetic units, sentence context and neighboring words affect the pro; nunciation of individual lexical items.

The goal of speech perception is to overcome the variability associ-, ated with actual speech, segment the speech signal, and identify tht individual phonetic segments that underlie it. Despite the extreme vari­ability of the speech signal, various acoustic cues are associated with distinct phonetic units. Vowels, glides (such as /w/ and /y/), and li< uids (/r/ and /1/) have characteristic configurations of formants. Stof are usually associated with a few milliseconds of silence, followed by burst and characteristic formant transition into the following vowel, glide, or liquid. Nasals (/n/ and /m/) have a distinctive acoustic fea­ture, whereas fricatives are noisy. While these are by no means invari­able cues, they certainly provide a great deal of information. Since the hearer is also a speaker, he can compensate for much of the variability produced by speaker characteristics, like speech rate and shouting. People seem to use knowledge of their speech production in their per­ception of the speech of others.

Psycholinguists can say two things about the process of speech perception with confidence. First, the speech process is possible only because people know a language that contains phonemes. Consider the dramatic variability in the acoustic signal representing /d/ in Fig­ure 5.7. By observing the formant transitions, it is clear that they do not look anything alike. Nor do they sound anything alike if one eliminates the vowel. Actually, it is not possible to have a stop consonant without a vowel. If one took /di/, /da/, and /du/ and made the vowel shorter and shorter (it is possible to do this instrumentally) until it went away, one would no longer hear a /d/. Instead, one would hear little chirps, rising in the case of /di/, falling in the case of /du/. These do not sound at all alike. So where is the /d/? It is literally in the hearer's mind, and not in the physical signal. The speech signals associated with

i n r n r W P I >|I PPm i i

llii' throe syllables contnln tin* information that allows one to recover tin1 throe /d/s that wore in the mind of the speaker. However, the phys­ical signal does not contain three identical acoustic events.

The phenomenon of categorical perception (Abramson «St Lisker, IT/O) illustrates the powerful effect that one's phonemic system has on *»|M‘irh perception. To illustrate this, the categorical perception of the voicing distinction in stop consonants needs to be described. The pri- iiiiiry acoustic difference between the voiceless /1/ and the voiced /d/ In simple syllables like "ti" and "di" is the time that elapses between i vloase of the closure between the tongue tip and the alveolar ridge and tin* onset of voicing for the following vowel. This is called voice onset lime (VOT). For /d/, voicing begins either the moment the closure is released (for a VOT of 0) or within the first 30 ms after release of the clo- Niire (VOTs of 1, 2, 3 . . . 30 ms). However, the /t/ has longer VOTs—I H‘tween 40 and 100 ms. This is an excellent example of variability in the Npoech signal. There are many different varieties of /d/ (with different VOTs) and there are many different varieties of /1/ (with different V( )Ts). These are not separate allophones of /d/ and /t/, because, lin­guistically, no significant difference exists among them. The basic fact about categorical perception is that people perceive all the /d/s as the same and cannot even discriminate among them, which holds true for the /t/s as well. However, an alveolar stop with a VOT of 20 ms will be perceived very differently from one with a VOT of 40 ms. The first will sound like a /d/, the latter like a /t/. Thus, there is a clear boundary between the VOTs of the physical signals representing /d/ and /t/.I Vople perceive a clear distinction between pairs of sounds drawn from each of those categories, but they perceptually ignore variability within the categories.

The following is a typical kind of experiment demonstrating the phenomenon of categorical perception (Yeni-Komshian, 1993). A speech scientist (with a computer) creates a series of /di/ and /ti/ syl­lables which vary only in VOT. There are VOT differences of 10 ms, which result in seven different acoustic signals, varying from a VOT of0 to one of 60. People listen to these syllables and are asked what they heard. The graph in Figure 6.3 illustrates an idealized categorical per­ception curve. The line joined by little o 's indicates the percentage of subjects who perceive the signal as /di/, while the line joined by little x's indicates the percentage of subjects who perceive the signal as /ti/. The syllables with 0,10, and 20 ms VOT are perceived by qyeryone as

O = /di/ X = /ti/

VOT (in msecs)

Figure 6.3. Atypical categorical perception curve.

/di/, while those with 40, 50, and 60 are all /ti/. The syllables with 30 ms VOT, which falls right at the boundary mark, is sometimes reported as /di/ and sometimes as /ti/. This is known as the cross-ova point. This illustrates the first aspect of categorical perception—physi­cally different acoustic signals are categorized by the perceptual system as belonging to the same phonemic category. The same people are then presented with pairs of syllables and asked if they are the same or dif­ferent. If the pairs are taken from different categories, subjects are right 100% of the time. The syllable with 20 ms VOT is perceived as com­pletely different from the one with 40 ms VOT. However, within the same category, the same physical difference (20 ms) does not produce a

l>eiveptual difference. Tin1 syllable witli 40 ms VOT is not discriminable Iroin the one with 60 ms VOT That this is a perceptual rather than an auditory phenomenon is suggested by an experiment in which sub­let Is' decisions about sameness were timed. It took subjects slightly longer to say that two dissimilar members of the same category were Ihe same (e.g., the syllables with 40 versus 60 ms VOT) than to say that I wo identical members were the same (e.g., two syllables with 40 msec V() I or two with 60 ms VOT) (Tash & Pisoni, 1973). This finding is compatible with the view that the auditory system registered the dif- I erence between the stimuli, but because the perceptual system recoded I lu*m as the same phoneme, they were not perceived as distinct.

VOT differences are not the only acoustic differences that produce categorical perception. It is also possible to manipulate formant transi- I ions so that a continuum of syllables can be created between /bi/ and /gi/, but they will be perceived categorically as /bi/, /di/, and /gi/. That this perceptual phenomenon is a property of the hearer's linguis- I ic system is demonstrated by cross-linguistic work in categorical per­ception. Subjects with different phonemic inventories perceive the same sequences very differently. For instance, an experiment with Thai speakers demonstrated that they perceived the /d/-/t/ continuum very differently from English speakers (Lisker & Abramson, 1964). This is because Thai has three categories of such stops. They not only have /d/ and /t/ as different phonemes, but they also have two distinct voiceless phonemes, an aspirated and an unaspirated /t/. Thus, Thai subjects divide the VOT continuum into three perceptual categories, while the English speakers divide it into only two.

The second thing psycholinguists can say with some confidence about the speech perception system is that it is constructive. This means that the speech perception system takes information anywhere it can find it and constructs a linguistic perception of the acoustic signal it hears. It has been demonstrated that a wealth of acoustic features are associated with the various types of phonetic elements. Another major source of information is the phonemic inventory of the hearer's lan­guage, as well as internalized information about how speech is pro­duced. The speech perception system can make adjustments and pre­dictions based on a speaker's vowel space and speech rate. Much of what psycholinguists know about the constructive nature of the speech perception system arises from the study of phonological illusions (much as psychologists have learned much about the visutl system by

studying optical illusions). One of these phonological illusions is t McGurk Effect (McGurk & MacDonald, I97H). This illusion illustrate the way hearers use visual information as well as auditory information in the construction of a phonological percept, and many versions of it exist. The following is a description of the way it works. An auditory recording is made of a person saying the syllables "bah," "bah," "bah/ "bah." Next, a video recording (without recording sound) is made of the same person saying "bee," "vee," "thee," "dee." The audio record« ing is then played in synchrony with the video of the person's mouth moving. What one perceives is "bah," "vah," "thah," "dah." It is al though your perceptual system takes the consonant from the visual stimulus and the vowel from the auditory stimulus, then puts them together so that you perceive syllables the person never actually said. It is a quite compelling illusion, but not really terribly surprising. Every­one has had the experience of being able to perceive speech better if the speaker is in view. Experiments have shown that if people are asked to report speech that has been made difficult to understand by embedding it in lots of background noise, the speech is much more easily under^ stood if the subjects can see the words being produced. Also, th lip-reading abilities of many deaf people are quite remarkable. This is another example of an earlier point, which is that knowledge of the way speech is produced is one of the kinds of information available to our speech perception system.

Another kind of illusion that illustrates the constructive nature of the speech perception system was created by Warren (1970) and is called phonemic restoration. An example of such an illusion is created if a person takes a sentence such as The state governors met with their respec­tive legislatures convening in the capital city. The person removes the mid­dle /s/ from the word legislatures and inserts a recording of a cough of exactly the same duration as the missing /s /. The person then plays the sentence to people and it sounds exactly like the /s/ has not been removed. Moreover, the people listening tend to hear the cough either before or after the word legislatures. This is a different phenomenon, known as perceptual displacement. If a stimulus arrives while a percep­tual unit is being processed, it will be perceived as occurring either before or after the perceptual unit. This phenomenon will be discussed in more detail in the next chapter. What is of interest here, however, is the fact that the perceptual system "fills in" the missing /s/. This is such a compelling illusion that people cannot even guess which

phoneme has been removed when (hey are told that one is missing. The Illusion does not work, however, if there is silence instead of the cough, lurlhermore, a pure tone does not produce a very good illusion. The production of /s/ involves making a hissing sound by passing air Ih rough an opening created by the tongue tip and teeth. The acoustic Hlect of this is similar to that of a cough, which is a burst of noise with­out formant structure. Thus, the illusion seems to work only if the replacement is similar acoustically to the removed segment. Another example of phonemic restoration can be achieved if one takes a wordIi ke slit and splices increasingly long periods of silence between the Is/ and the /1/. What one hears is /s-silence-lit/, /s-a little longer silence- lil/, and so on until the silent interval is about 30 ms. Then all of a sud­den one can hear "split" clear as a bell. Remember that when a stop consonant is produced there is initially complete silence because the vocal tract completely closes for a very brief period of time. That period is just about 30 ms. When the silence becomes longer, the "split" per­ception goes away. Note that any stop other than /p/ in that position would not produce a word of English (in fact, would violate the phono-tactic rules of English).

One explanation for these illusions involves a conception about how lexical retrieval works. The lexical retrieval system uses as much acoustic information as it has available to locate a word in the internal­ized lexicon. After the word has been retrieved, the full phonological representation of the word is then checked against what the person has heard. This is called post-access matching. If the match is goal enough, the word is accepted as correct and the full phonological representa­tion from the lexicon becomes the percept. This process allows even a degraded acoustic signal to provide enough information to allow retrieval to take place; the phonological details are then filled in by the phonological information associated with that lexical item. Taking this view into account, plenty of acoustic information was available in the above examples for the words legislatures and split to be accessed and to survive the post-access check. Thereafter, the /s/ and the /p/ were per­ceived based on the more complet^phonological information obtained from the lexicon, not on the initial acoustic information.

The fact that people can perceive the phonetic structure of nonsense words (e.g., plit) demonstrates that speech perception based solely on the acoustic signal is possible, with no assistance from the lexicon (because nonsense words are by definition not stored in tfce lexicon).

The existence of perceptual illusions like phonemic restoration, how« ever, demonstrates how the perceptual system can cope when it encoun­ters inadequate amounts of acoustic information.

Bottom-Up and Top-Down Information

A concept that is of great importance in psycholinguistics is the distinc­tion between bottom-up and top-down processing. Because psycholin- guistic processes are fundamentally information processing routines, it is easy to understand this distinction in terms of the kind of inform% tion used by the comprehension processes of the hearer. The following is an illustration of these processes. If I say a word to you in isolation, clearly and distinctly, you will be able to decode my acoustic signal and retrieve the word from your lexicon. So if I walk up to you and say, "cat. food," you might think I am a bit odd, but you would be able to retries* those words from your lexicon. In this situation you would be using only bottom-up information. To achieve lexical retrieval you have used only the information associated with a representation (in this case acoustic) as input to the decoding operation you are performing (in this case lexical retrieval). Suppose, on the other hand, we live together with a cat and you are going shopping. We are in different rooms, the disl r̂ washer is running noisily, and I call out to you a speech signal that badly degraded because of distance, background noise, and lack o§> visual information: "Fluffy's bowl is empty. Be sure to buy some cat food." You do not have good enough acoustic information to recover all the words I have said, but you catch "Fluffy" and either "bowl" or "empty," and then "buy." It is a good guess that "cat food" is in the sen­tence. You have understood this version of "cat food" by using top- down information. This is information that is not part of the acoustic representation of "cat food," but is, instead, contextual information that allows you to select which of the many things I might have been saying, given inadequate bottom-up information. In this case, part of the con­text was in the sentence— the words you did catch, especially the cat's name. Part of the context was also in the general situation. It is com­monplace for you to shop and buy Fluffy's food; I knew you were going shopping, and you knew that I knew. All of this conspires to allow you to understand "cat food" as the appropriate member of the set of things

11 ir* I Imi H S|ir>rM I i I Vm r| illi im iiiii 11 t*xl( ill At ( <*sn l / J

I might have been saying. Quito a bit ol this sort of thing probably goes mi in ordinary conversation. Recall from Chapter 3 that people with Hroca's aphasia are good at understanding conversational speech but poor at understanding sentences for which they had to do a detailed iin.ilysis. The suggestion is that they are using contextual (top-down) information to understand what is said to them. In one experiment, words were excised from sentence contexts and played to listeners (I ’ollack & Pickett, 1964). The listeners were very poor at understanding I he words, but when they heard the entire sentences they understood I he words without difficulty. The context of the sentences in which the words appeared, and their general meaning and structure, allowed the words to be understood despite inadequate bottom-up information. This is a typical example of the use of top-down information. When the bottom-up information inadequately specifies a word or phrase, top- down information can allow the hearer to select among a range of pos­sibilities. If bottom-up information is adequate, however, top-downinformation will not be necessary.

One experiment illustrates everything that has been said about bot­tom-up versus top-down information (Games & Bond, 1976). The experimenter created a series of stimuli that ranged from bait to date to gate. Remember that a continuum can be created by manipulating the formant transitions of stop consonants so that they will move from /b/ to /d/ to /g/. This is what was done in the experiment. There were "perfect" versions of /b/, /d/, and /g/, and there were other versions that were right in between /b/ and /d/ and /d/ and /g/ (analogous to the cross-over point in the VOT continuum in Figure 6.3, which is ambiguous between a /t/ and a /d/). In the experiment, there were acoustically perfect versions of each of the three words and acoustically ambiguous versions of each. They were embedded in the following sentence contexts (called carrier phrases):

Here's the fishing gear and th e________

Check the time and the_________

Paint the fence and the________

Clearly, the most plausible continuation for the first carrier phrase is bait, for the second, date, and for the third, gate. The subject's task was

simply to report which word had appeared in I hi* final position o f th# sentence. What do you think happened? When the acoustic signal wi| absolutely clear—that is, the perfect versions of the three words—they were correctly perceived even if they appeared with a less plausible carrier phrase. If the acoustic signal was ambiguous, however, the word was reported as compatible with the carrier phrase. Consider the per­fect version of date. It would be correctly reported as "date" even if It appeared in the sentence Here's the fishing gear and the date. The ambi§u> ous signal, which was halfway between "bait" and "date," ("b / dat^ was heard as "bait" if it appeared in "Here's the fishing gear and-4i|f b/date," but as "date" if it appeared in the sentence "Check the tiq^ and the b/date." The reason this experiment is a good example is that it shows that the hearer will depend completely on bottom-up informa­tion if it is clear and unambiguous even if it produces an implausible outcome. If the stimulus is ambiguous and does not produce a unique percept, then top-down information will be recruited, and the ambigu­ity will be resolved in favor of general plausibility considerations. In this experiment, the flip side of categorical perception is observed. The classical categorical perception experiment described earlier demon­strates that different acoustic signals can be perceived as belonging to the same phonemic category. The Games and Bond study demonstrates that the same acoustic signal can be perceived as different phonemes under certain circumstances. More recently, Borsky, Tuller, and Shapiro (1998) have also demonstrated this phenomenon.

Lexical Retrieval and the Hearer

The speaker enters the lexicon using information about meaning so she can retrieve the phonological structure of the appropriate words to con­vey the meaning she is constructing for a sentence. The speaker finds a lexical entry that matches her meaning representation and then the pho­nological information is available to her. The hearer's task is the oppo­site. It is to use the phonological representation (decoded from the acoustic signal) to retrieve information about meaning. The hearer finds a lexical entry in which the phonological representation matches the one he has heard. Next, the word is retrieved, and the information about its meaning and structure is then available to him. As discussed earlier, the speed at which lexical retrieval takes place remains a mys-

I h p H e f l i e i I ’PI l F | I I K i h . m u I ' " "

i«ry. It occurs rapidly and unconsciously, yet each time it happens, the Imarer must consult a lexicon of at least 80,000 items.

One of the techniques psycholinguists use to investigate the proper- lica of words that influence their ease of retrieval is called a lexical deci- tion task. In this task the subject sees a string of letters and must push a billion marked yes if the letters constitute a word of English and no if I hey do not. The responses are timed, and the question in this task is what kinds of letter strings are responded to more rapidly and which ones are responded to m ore slowly. Responses on such a task are very rapid, in the hundreds of milliseconds (much less than a second), and I lie measurable differences are very small. However, the responses are statistically reliable and give psycholinguists a great deal of informa­tion about the relative speed of retrieval of words with different prop­erties. The following are 14 letter strings. Beside each write either yes or no, indicating how you would respond if you were a subject in a lexical decision experiment. Next, see if you can figure out what properties of the words might affect the speed with which you would make yourjudgment.

clock skern bank doctor nurse tlat plim

You probably responded no to four of the letter strings that are not real English words. What is interesting is that in a lexical decision task you would have responded to two of them very quickly and to the other two very slowly. Zner and tlat are not possible words in English. They violate the phonotactic rules of the language. Such nonwords are rejected very rapidly. It is as though the lexical retrieval system carries out some sort of phonological screening, and, if the string is not a pos­sible word, does not even bother looking in the lexicon. The two possi­ble nonwords, skern and plim, however, would take much longer to reject, as though the retrieval system keeps looking but never finds an entry. Bower (1990) reported that a neurologist named Petersen demon­strated just this kind of brain activity using a technique called positron

bookhuttablechairurnbatzner

\ ?(> , * |\y( holintjiilslii s

emission tomography (PET), which measure» blood-tlow changes in brain. This technique is able to show the location of activity in the br§|p by color-coded images generated by a computer. Petersen presented his subjects with impossible nonwords, possible nonwords, and real word^ He found that there is one area of the brain that responds to possible and real words (but not impossible nonwords), and yet another a m that responds only to real words. Apparently, there is a brain system that recognizes phonotactically possible words, screening out im p o s t ble ones; Another area distinguishes between possible nonwosds and real words. Both of these areas are in the left hemisphere of the brain only.

Of the real words, the more frequent ones, such as book and c lod ^ would be responded to much faster than low frequency ones, such as hut and urn. It is not that people make errors with the rare words (although some subjects might do so), it is that the yes responses w it^ these words are slower than the yes responses to more common words (Rubenstein, Garfield, & Millikan, 1970). This suggests that words that are used often are somehow more available to the lexical retrieval system.

Another property of words that makes them more accessible in a lexical decision task is ambiguity. The more different meanings a word has, the faster its retrieval time. It is easy to see why this would be the case. Assume that each meaning of a lexical ambiguity constitutes a separate lexical entry, so that bank would have one entry for river bank and another for money bank, and bat would have one entry for the flying mammal and another for a piece of sports equipment. Suppose there is a bowl full of white marbles with a few black marbles in it. If one bowl has one black marble and the other has two, a person is more likely, searching randomly, to come upon a black marble in the bowl with two, than in the bowl with just one. Moreover, the more black mar­bles in the bowl, the faster a person will find one. This analogy makes it easy to understand the rapid access for ambiguous words in relation to unambiguous ones. The analogy also supports the idea that these is a separate lexical entry for each meaning of an ambiguous word. Such words are also called homophones, meaning that there are two sep­arate words which have identical phonological representations. Some homophones, like bat, are also spelled the same in both versions; other homophones are not spelled the same, such as their and there. Moreover, others are homophonous in some dialect areas, but not in others. For

Ih p I lea iM S|it*rc h I ’m rp n o ii .iiu l I rxl< .il A( ( rss I / /

instance, in some dialirts of American English, marry and merry are homophones; whereas in others they are not.

The last characteristic of a word that speeds its retrieval time is ( .tiled priming (Meyer & Schvaneveldt, 1971). Actually, this is not a property of a word, but a reflection of a process that is very general and interesting. In your list you have two word pairs, doctor-nurse and liiblc-chair. In each case the two words are close associates of one another. In a word association test, nurse would likely be a response to iloctor, and chair to table. When a word is retrieved from the internal lex­icon, for a brief period of time (milliseconds), words that are closely associated with that word will become more available for retrieval. Thus, if retrieval of such a word is required, it will take place faster in that small window of time than under ordinary circumstances. This is an example of priming. We say, then, that when doctor is retrieved, it primes nurse and other related words, such as maybe scalpel, stethoscope, hospital, and so on. When nurse is the next item in the lexical decision, it will be responded to a few milliseconds faster than it would be if it were preceded by an unrelated word, such as car. Similarly, table would prime chair. Other similarities between words can also result in prim­ing. For instance, words that rhyme prime one another. The process of priming is probably not limited to psycholinguistic materials, nor is it exclusively related to lexical access. Pictures could undoubtedly be shown to prime other pictures of related scenes. Association is a cogni­tive, rather than a purely linguistic or psycholinguistic, phenomenon. Concepts and ideas are associatively related, just as words are. Thus, while psycholinguists have investigated priming, it is not unique to the language module of the brain. How psycholinguists have used priming as a research tool will be discussed later.

An interesting and unresolved issue in lexical access is the role of bound morphemes. If stems are stored in the lexicon and derived forms are not, then the bound morphemes of the derived forms must be removed before lexical access can take place. This is called morpheme stripping (Taft, 1981). It is almost certain that regular inflectional mor­phemes, like the regular past tense marker, spelled -ed, are stripped. People almost certainly do not store walked in their lexicon as well as walk. When those items are encountered, the bound morpheme is removed and the stem is accessed. But what about derivational mor­phemes? These cases have elicited a great deal of debate among psy­cholinguists. Derivational morphemes vary in their productivity. The

I SH , » I ’syi lin lin ijulslk s

agentive -er as in rider, player, singer, and hi» on, in highly productive and can be attached to any verb indicating an activity. It would seem that economy would be served if the derived forms of words created from highly productive derivational morphemes were not stored in the lexi­con, but instead were created for production and decomposed for retrieval. Indeed, Bradley (1980) produces elegant and convincing evi­dence that the derived forms of stems plus -er, -ness, and -ment are not stored. These are highly productive morphemes that do not alter the pronunciation of the stems to which they are attached. In Bradley's study, derived forms containing a less productive derivational mor­pheme -tion (as in derivation), which does alter the pronunciation of the stem to which it is attached, were hypothesized to be stored in the lexi­con. There are many derivational morphemes yet to be investigated, but the one- thing psycholinguists can tentatively conclude at this time is that some are stored with their stems and some are stripped for access. Obviously, the size of one's lexicon depends on how many derived forms are stored separately from their stems. If many derived forms are listed separately from their stems, then the lexicon will be larger than a lexicon in which only stems are stored.

The Cohort Theory of Lexical Access

It is important to understand more about the rapid and unconscious re­trieval of words from the lexicon. A contemporary theory that accounts for many facts about lexical retrieval and has generated a great deal of productive research is called cohort theory. Cohort theory was originally described by Marslen-Wilson and his colleagues (Marslen-Wilson, 1987; Marslen-Wilson & Tyler, 1980). According to this theory, acoustic infor­mation is rapidly transformed into phonological information, and lexi­cal entries that match phonologically are activated. After the first syl­lable of a word is received, a large number of lexical entries will be activated; after the second syllable is received, a subset of those will remain activated (when an entry ceases to match, it deactivates). Finally, at some point—before the end of the word if it is unambigu­ous—a single lexical entry will be uniquely specified, and it will be retrieved. This is called the recognition point of a word, and on average it occurs within 200 to 250 ms of the beginning of the word. The fact that

f | )r | irvuri >,|ippi 11 I'pii r i Mini i , it k11 exit ill At n ’ss . * I M

words can be retrieved before they are completed has been dem­onstrated by Holcomb and Neville (1991) in an experiment examining brain waves associated with sentence processing. Recall from Chap­ter 3 that there is a characteristic wave, labeled N400, that is associated with the presence in a sentence of a semantically anomalous word. Hol- iumb and Neville (1991) showed that this effect begins long before the entirety of the semantically anomalous word has been encountered by Ihe listener. According to this theory, an initial cohort of phonologically similar words are activated, and at its recognition point one is selected .ind integrated into the representation of the sentence being constructed.I f this results in a semantic anomaly, an N400 wave is the electrophysi- ological result.

The cohort theory of lexical access predicts that the initial part of a word will be more important for lexical access than its end, a prediction that has been confirmed by a number of different kinds of experiments. Mispronunciations at the beginnings of words are detected more accu­rately than are mispronunciations at the ends of words (Cole, Jakimik, & Cooper, 1978). In a variation of the phoneme restoration phenome­non, subjects were found to restore phonemes at the ends of words more often than phonemes at the beginnings of words (Marslen-Wilson & Welsh, 1978). Both these findings suggest that the main work of lexi­cal retrieval was done by the first part of the word; by the time the end of the word was received, the lexical item had already been retrieved, and thus little attention was being paid to the actual acoustic signal associated with the end of the word.

This lexical retrieval theory (as well as many other similar ones) assumes that every word in the hearer's lexicon has some resting level of activation. Stimulation by matching phonological information increases a word's level of activation. When activation reaches some threshold level, the word is retrieved and is then available for use in the ongoing sentence processing routine (or in a lexical decision task). The notion of activation can also account for the observed frequency effects in lexical retrieval. The explanation is that high-frequency words have a higher resting level of activation than do low-frequency words. Since retrieval depends on a lexical item reaching some threshold of activa­tion, high-frequency words will reach that threshold faster than low- frequency words. Ambiguous words are just those for which there is no single phonological match available, and thus all representations reach$

Remembering Sentences, Processing Texts, and

Having Conversations

Psycholinguists are interested in two very different kinds of memory, both of which are involved in processing sentences and using language. Working memory is the temporary stor­

age space for phonetic information as sentences are being processed. Long-term memory is critical when those sentences are being put to use in conversations and in processing text.

Working Memory and Sentence Processing

While a sentence is being processed, words are held in working memory. This is a storage system where information is retained for very brief periods of time before it is sent on in a recoded form to longer-term memory. The recoded information is sometimes called chunks, a term coined by the famous psycholinguist George Miller (1956). He discov­ered that people can remember approximately 5 to 9 chunks of infor­mation. In an article entitled "The Magical Number Seven Plus or Minus Two," Miller reported experiments demonstrating that people can recall 5 to 9 single letters, short words, or short sentences, each of which constitute 5 to 9 chunks of information. The task of #ie sentence

171

comprehension system, then, is lo recode lexical anti structural info» mation to create a representation o f sentence meaning, which is then sent to longer-term memory.

The role of working memory in sentence processing has been inves­tigated in a number of ways. Early studies (Foss & Cairns, 1970; Wan­ner & Maratsos, 1978) showed that requiring subjects to recall word lists impaired their ability to process sentences. Furthermore, sentences of greater processing complexity, such as those with object relativi clauses (e.g., The reporter that the senator attacked admitted the error), were more adversely affected than less complex sentences, such as those with the analogous subject relative (The reporter that attacked the senator admitted the error.). This is an anticipated result if memory resources are required to process a sentence. Also, the more difficult a sentence, the more resources are needed to process it.

More recently, researchers have investigated the processing abilities of people with varying working memory spans. For instance, Just and Carpenter (1992) showed that people with low memory spans have more difficulty with object relative clauses than do people with high spans. This is true presumably because in order to understand such a sentence, a person must maintain the reporter in working memory so it will be available to fill the gap following attacked. The better the mem­ory span, the more available the filler will be. Carpenter, Miyake, and Just (1994) also reported studies showing that following access of all the meanings of an ambiguous lexical item, people with high memory spans were able to maintain them in working memory for longer peri­ods of time than people with more limited spans. As a result, high-span people are less adversely affected by lengthy delay of disambiguating information than are low-span people. Studies also report that low- span, but not high-span, people show garden path effects for sentences such as The evidence examined by the lawyer shocked the jury (Ferreira & Clifton, 1986). They suggest that high-span people are able to hold structures in working memory long enough for nonlinguistic informa­tion to affect processing, while low-span people are not. Memory span limitations are also implicated, they propose, in the language process­ing difficulties of people with aphasia and the elderly.

While working memory is involved in the immediate processing of individual sentences, long-term memory is involved in all language use on a daily basis. Individual sentences are processed rapidly, but they must be retained and integrated with other information if people are to

l\Vmi'ml>t'iln<| Srnlr'iii M, I’Miinwlnu <in<l I l.ivlmi ( onvm.wons . * i / ■>

use them in conversation or in the interpretation of any ongoing narra­tive, such as a play, a story, or a film. Thus, short-term memory is asso­ciated with obtaining the basic building blocks of sentence meanings. I ,ong-term memory is crucial for actually putting those sentence mean­ings to use.

Memory for Sentences

Three important things happen to sentences in long-term memory. First, information about structure and even individual lexical items is lost, while meaning is retained. Second, meanings of many sentences are combined, so individual sentences no longer have separate repre­sentations. Finally, inferences are added to representations of meaning.

A well-known early experiment demonstrated that information about form (structure) is not retained in memory, but information about meaning (content) is (Sachs, 1967). Subjects listened to a narrative, which contained the sentence He sent a letter about it to Galileo, the great Italian scientist. The subjects then heard a probe sentence and were asked if they had heard it in the narrative. The logic of the experiment was that when the subject was asked whether she had heard the sen­tence, she would search her memory representation of the narrative she had heard, attempting to match the probe sentence with part of that rep­resentation. If there was a mismatch between the probe sentence and the memory representation, she would report that she had not heard the sentence in the narrative; if it did match, she would report having heard it. The probe sentence was either the same as one the subjects had heard, or it was changed in its form, as in He sent Galileo, the great Italian scientist, a letter about it; or the meaning of the sentence was changed, as in Galileo, the great Italian scientist, sent him a letter about it. Immediately after hearing the sentence, subjects could recognize both kinds of changes. However, after a brief period of time (less than a minute), they could recognize if the sentence had been changed in meaning, but not if it had been changed in form only. This result demonstrates that the subjects had retained a representation of the meaning of the sentences they had heard, but not the exact form. It is a perfectly reasonable find­ing, since sentence structure exists only to determine the basic meaning of a sentence. Once that task has been accomplished, it is completely useless and there is no reason to store it in memory. It is dPvery general

phenomenon that people recall the gist (general meaning) of what they have heard, but not the surface form. People who are bilingual and listen to radio stations in both their languages report that they often remember the content of a particular news report but do not recall which language they originally heard it in. It is very difficult to remem­ber the precise form of what anyone has said. Sentences that carry so- called high interactive content, such as Do you always put your foot in your mouth? or Can't you do anything right? are exceptions to this general principle (Keenan, MacWhinney, & Mayhew, 1977). The exact form of sentences with great interpersonal import are more likely to be recalled than neutral ones.

Another reason it is difficult to remember exactly what was said is because individual sentence meanings are integrated to create more global representations of meaning. An experiment has demonstrated this phenomenon. Experimenters (Bransford, Barclay, & Franks, 1972) communicated four pieces of information: Ants were in the kitchen; ants ate jelly; the jelly was sweet; the jelly was on the table. To convey this information, Bransford et al. created three kinds of sentences. One sentence contained one piece of information, such as The ants are in the kitchen; another type contained two pieces of information, such as The ants in the kitchen ate the jelly; and the third type conveyed three of the pieces of information, such as The ants in the kitchen ate the sweet jelly or The sweet jelly was on the table. Thus, all four pieces of information were contained in various sentences containing one, two, or three pieces of information. All four pieces of information were never contained in one sentence. The experimenters presented these sentences, followed by a recall task. Subjects could not distinguish between sentences they had actually heard and different sentences containing the same information in different configurations. The most interesting finding was that the subjects reported, with a high degree of confidence, that they had heard the completely integrated sentence containing four pieces of informa­tion, The ants in the kitchen ate the sweet jelly which was on the table, which had never been presented. This experiment illustrates that the memory system is very good at integrating and synthesizing information. How­ever, it is not good at keeping individual bits of information distinct from others.

Unlike the sentence processing system, the memory system is decidedly not modular. It recruits real-world knowledge, makes infer­

K’ciTii'iMl)('iln(| SpmIpiiifv MinimMm| Irxiv .if»111.ivln<| ( onvm.illnn\ I f'*

ences, and stores these inferences right along with the information it actually received. Johnson, Bransford, and Solomon (1973) demon­strated that subjects would make and store in memory inferences about objects never mentioned in sentences they heard. In this study, the experimenters read identical paragraphs to subjects, except that one paragraph contained the clause He was pounding the nail when and the other contained He was looking for the nail when. Subjects who heard the first version were likely to believe they had heard the word hammer in the paragraph, but the others were sure they had not. Using real-world knowledge about the usual method of pounding nails, the subjects who heard the first version made the inference that a hammer was being used and stored that information in their memory representation of the sentence. This type of inference is called an instrumental inference. An experiment by Bransford, Barclay, and Franks (1972) demonstrated that people add spatial inferences to their memory representation of sen­tences. Subjects heard a sentence such as Three turtles sat on a floating log and a fish swam beneath it. Only 3 seconds later they could not report whether they had heard that sentence or Three turtles sat on a floating log and a fish swam beneath them. This is because, given their knowledge of spatial relationships, if the fish swims beneath the log and the turtles are on the log, then by inference the fish has swum beneath the turtles as well. Another group of subjects, who heard Three turtles sat beside a float- ing log and a fish swam beneath it, could easily report that they had not heard Three turtles sat beside a floating log and a fish swam beneath them, because beside does not allow the same spatial inference that on does. This illustrates another reason why it is virtually impossible for people to recall exactly what has been said to them. Not only do people inte­grate a wide variety of meanings, they add to their memory all the inferences they have made at the time they originally heard the speech. Note that inferences are not part of the basic sentence meaning con­structed by the sentence processor. This was demonstrated in a study by Jenkins (1971). Following the two types of turtle sentences (on vs. beside), he asked subjects, "Did the fish swim beneath the turtles?" The correct answer in either case was yes, but the answer is based on the basic meaning of Three turtles rested on a floating log and a fish swam beneath them, and on the inferred meaning of Three turtles rested on a floating log and a fish swam beneath it. It took subjects significantly longerto answer yes when they had heard the second sentence than when theyI*

had heard the first. Information that had lo In* inferred was, therefor#, not available as quickly as information contained in the basic meaning of the sentence. Singer (1979) reported a similar finding.

All these facts about sentence memory should have important entailments for so-called "eye witness" testimony— or in this case "ear witness" testimony. A psychologist a the University of Washington, Elizabeth Loftus, has spent her career investigating aspects of memory that may have legal implications. One of her experiments (Loftus tt Palmer, 1974) demonstrated not only the role of inferences in memory but also the endless possibilities for manipulating the memories of oth­ers. Two groups of subjects saw a film of an automobile accident. One group was asked, "How fast was the Buick going when it hit the Ford?" The subjects estimated 34 mph. The other group was asked, "How fast was the Buick going when it smashed into the Ford?" They estimated 41 mph, a statistically significant difference. A week later the subjects were asked whether there had been any broken glass at the scene of the accident (there had been none). Only 14% of the people who were ini­tially asked the question using the word hit said there was broken glass, but 32% of the people who heard the question with smashed into erro­neously reported broken glass. The group who heard smashed had been led to store in memory a more violent representation of the original scene than did the other group.

There is a famous story about the Swiss psychologist Jean Piaget. He grew up believing that his earliest childhood memory was of being rescued from kidnappers by his nanny. Years later he discovered that she had made up the story to curry favor with his parents. His was completely a constructed memory, based on a story that he had been told repeatedly (Piaget, 1962, pp. 187-188). If you want to find out about someone's earliest memories, you should be careful not to rely on memories of events they have been told about, seen pictures of, and so on. Human memory is not a simple recording device. It is a complex, dynamic system that constructs memories based on many factors, only one of which is what was actually experienced.

Memory for sentences is not really about sentences. It is about the acquisition of information by the use of language. Memory for sen­tences is no different from memory for other kinds of things. The same principles— losing form and retaining meaning, integrating across a variety of experiences, and adding inferences based on knowledge from a variety of sources— apply to memory for all kinds of events.

Text Processing

An organized sequence of connected sentences is called a text. Stories, narratives, and discourse, written or spoken, are all examples of texts. When a person understands a text, he must take the basic meanings of the individual sentences and integrate them into a semantically coher­ent framework.' In order to do this, links must be discovered between and among the sentences of the text. These links are both semantic and referential. The goal of text integration requires at least two major pro­cesses, anaphoric reference and inference. Although these two processes interact in many cases, they will be discussed separately.

Anaphoric Reference

An anaphor is a linguistic device that refers to someone or something that has been mentioned in the previous context. It can be either a pro­noun, as in John came home yesterday. He is on Spring Break, or a noun phrase introduced by a definite article, as in I got a new kitten yesterday. The little darling slept with me last night. He and the little darling are both anaphors; in order to understand the pairs of sentences, the anaphors must be matched with their referents, John and a new kitten, respectively. Pronouns are anaphoric because they cannot be interpreted without locating an antecedent for them; they have no independent meaning, except for gender, number, and case, which is given by their grammat­ical form. As discussed in the previous chapter, there are grammatical principles that restrict the referents of pronouns, but there are no gram­matical principles for locating their referents among the set of gram­matically possible ones. Pronominal reference is an aspect of linguistic performance, unlike sentence processing, for which all kinds of nonlin-guistic knowledge is recruited.

Definite noun phrases are anaphoric for a different reason. They are anaphoric because the use of the definite article the presupposes that the referent of that noun phrase is already in the discourse. The first mention of an entity is introduced with an indefinite article (as in a new kitten above). Later reference to the same entity requires the use of the definite article. The entailment for the hearer is that when a noun phrase with a definite article is encountered it must refer back to an ear­lier instantiation of the same referent— it is anaphoric. (An exception to

this is the use of the definite article U* refer lo a spec ies, as in I'hc tiger isa carnivore.) This section will explore some of the factors that influencea person's ability to locate referents for the anaphors they encounter in texts.

The use of an anaphor in a text has two purposes. First, it is a cohe­sive device, anchoring a sentence to prior representations in the text. Second, the anaphor is a device for creating a semantically coherent text, and the referent of a pronoun will be resolved in such a way as to pro­duce the most semantically plausible meaning possible. For example, the pronouns in the second sentence of the following pairs of sentences are resolved very differently: Bill wanted to lend his friend some money. He was hard up and really needed it and Bill wanted to lend his friend some money. However, he was hard up and couldn't afford to (Garrod & Sanford, 1994). The sentences above both contain an ambiguous pronoun, as both Bill and his friend can be construed as male; the referent of he is determined semantically, referring to his friend in the first sentence and to Bill in the second. It is also the case that reference for even unam­biguous pronouns is facilitated by the overall plausibility of the result­ing interpretation. Stevenson and Vitkovitch (1986) demonstrated that the assignment of Henry as the antecedent of he is faster for the sentence Henry jumped across the ravine and he fell into the river than for the sen­tence Henry jumped across the ravine and he picked up some money. Although he is completely unambiguous in both cases, it is easier to imagine falling into the river than picking up money as being a conse­quence of jumping across a ravine. Therefore, identifying he as Henry produces a more semantically coherent interpretation in the first sen­tence than in the second.

Another important factor in the assignment of anaphoric reference is discourse focus. In general, a referent is more available if it is focused, and there are many ways this can be accomplished. Recency is one way to achieve focus; in general, more recently mentioned characters will be in focus, so near referents will be located more quickly than will more distant referents. In an eye tracking study, Ehrlich and Rayner (1983) demonstrated that when an antecedent was distant, readers spent a longer amount of time fixating the region immediately after the pro­noun than when an antecedent was near. Clark and Sengul (1979) showed that the immediately prior sentence has a privileged status in terms of availability of a referent for an anaphor. They took three- sentence sets, such as the following:

KViiirmbfilnii S m lm tn , Mini r«ln<| It'xlv .iml I l.ivm<| < onvrrs.illons . * 11)

A broad loom rug in rose and purple colors covered the floor.

Dim light from a small brass lamp cast shadows on the walls.

In one corner of the room was an upholstered chair.

Subjects read the three sentences, which were presented in varying orders. They were then given a target sentence and asked to push a but­ton when they had understood it. The target sentence following the above set was The chair appeared to be an antique. The chair was, of course, an anaphor whose referent was an upholstered chair. Reading times were much faster when the referent was in the third sentence than when it was in either the first or the second sentence. There was a much smaller difference between the availability of the referent from sentence one to sentence two than between either sentence and sentence three. A sub­sequent experiment demonstrated the same effect even if sentences two and three were joined as two clauses in a sentence. Therefore, the clause preceding the anaphor is privileged for referent location, not the entire sentence.

Other factors contribute to long-term focusing throughout a text. If a character is a main character, he is more likely to be in focus and more available for anaphoric reference. A character who has been introduced by a proper name is more likely to be a main character and in focus than one who has been introduced by a role description. Consider the fol­lowing pairs of sentences: The priest waited by the church wall. Mary walked up to him and Father Brown waited by the church wall. The woman walked up to him. In a reading time study (Sanford, Moar, & Garrod, 1988), exper­imenters demonstrated that sentences containing references to named characters—Mary in the first pair, Father Brown in the second— are read more rapidly than those containing references to characters who are notnamed.

Another focusing device is position in a prior sentence, with the subject being in focus position. Antecedent location is facilitated if the pronoun refers to the subject of a previous sentence. Hudson, Tanen- haus, and Dell (1986) compared reading time for sentences with pro­nouns of the following type: Jack apologized profusely to Josh. He had been rude to Josh yesterday versus Jack apologized profusely to Josh. He had been offended by Jack's comments. The former was read faster than the lat­ter because he refers to Josh, the referent in subject position and thus in focus.

Independent of discourse focus, Matthews (I486) showed th» when antecedents and pronouns are in the same complex sentence, t . _ structure of the sentence affects how quickly the antecedent can bt located. A more deeply embedded antecedent, such as the object of i prepositional phrase, takes measurably longer to locate than one that 1| closer to the surface, such as a subject or object noun phrase.

Locating referents for anaphors is essential to building a connected, semantically coherent mental representation of a sequence of sentence! that makes up a text. The more available a referent, in terms of recency and general importance in the text, the greater the facility with which it will be discovered and integrated into one's ongoing representation of the text.

Making Inferences

Memory for sentences (and just about everything else) is enhanced by stored inferences, which are indistinguishable in memory from the sen­tences that were actually experienced. The formation and storage of inferences is also a central feature of text processing. Even the shortest stretches of text require the reader to make inferences in order to connect the sentences into a coherent structure. Roger Schank (1976) has called inference the "core of the understanding process" (p. 168). Inferences are involved in the location of referents for anaphors. Haviland and Clark (1974) demonstrated in numerous studies the existence of bridging infer­ences connecting sentences in a text. This type of inference is illustrated by the following two sentences: We checked the picnic supplies. The beer was warm. The coherent interpretation of this pair of sentences requires the hearer to infer that the beer was part of the picnic supplies. Processing is facilitated or impaired depending upon the ease with which hearers will be able to make such inferences. The bridging inference above is easier than the one required to find the referent for the fire in A careless tourist threw a lighted match out o f his car window. The fire destroyed several acres o f virgin forest. Even more difficult is the bridging inference required for the referent of the woman in We went to a wedding. The woman wore white. These two sentences illustrate that if too many bridging inferences are required, the discourse can sound decidedly odd.

Inferences can do more than locate referents for definite noun phrases. They can also enhance their meaning. Consider the meaning of

•i noun phrase such as Hit* container, which is a very genera] term, with a nonspecific meaning. When used in a sentence, however, it can take on more specific meaning. For example, contrast the container held the soup with the container held the gas. By inference, the former is a small bowl or cup whereas the latter is a closed metal cylinder. Anderson et al. (1976) called this increased specificity of meaning the instantiation of general terms. These researchers demonstrated that instantiation does take place and that the more specific meaning is stored in memory. In this experiment, subjects heard a list of sentences, which included either The woman was outstanding in the theater or The woman worked near the theater. Their memory was probed using either the woman or the actress. The woman was an equally good memory probe for both sentences, whereas the actress was no better for subjects who had heard the second sen­tence. However, the actress as a probe enabled the subjects who had heard the first sentence to recall it twice as often as they did following the probe the woman. This finding is particularly important because it illustrates that instantiation is not simply the result of a simple associa­tion between, in this case, the woman and the theater. The inferences that allowed the actress (which had not even appeared in the sentence) to be a good memory probe were very specific to the subjects' real-world knowledge about the relationship that likely exists when a woman isoutstanding in the theater.

Bridging inferences are backward inferences in the sense that they require the hearer to review previous information in the text to provide coherence with a current item. For example, it is not until one encoun­ters the beer that one infers that the picnic supplies contained beer. For­ward or elaborative inferences are those made immediately after a piece of text is encountered, whether or not it is needed for coherence. If someone says, "John accidentally dropped his wine glass on the car­pet," and a person infers that the wine spilled, that is an elaborative inference. If someone says, "John accidentally dropped his wine glass on the stone patio," and one infers that the glass broke, that is also an elaborative inference. The inferences that lead to the instantiation of general terms as demonstrated above are examples of elaborative infer- ences. It is unclear under what circumstances people create elaborative inferences, but it is almost certain that they often do. Since elaborative inferences are not necessary for discourse coherence, they are not as vital to text comprehension as bridging inferences are. Singer (1994, p. 488) reported an experiment in which subjects were asketi to verify a

statement such as A dentist pulled tin' tm III after Inuring one of the fol­lowing three types of sentence pairs:

The dentist pulled the tooth painlessly. The patient liked the new method.

The tooth was pulled painlessly The dentist used a new method.

The tooth was pulled painlessly. The patient liked the new method.

In the first pair of sentences, no inferences are involved. The first sentence states explicitly that the dentist pulled the tooth. In the second pair, the bridging inference that the dentist pulled the tooth had to be made immediately in order to connect the two sentences in a semanti­cally coherent way. In the third pair, however, it is not necessary to cre­ate the bridging inference that the dentist pulled the tooth. If the infer­ence were made that the dentist was the puller of the tooth, it would have been an elaborative inference, not a necessary one. (The sentences are coherent with the inference that the patient was the possessor of the tooth, but that is irrelevant to this study.) The point is that in the first two pairs of sentences, one can assume that the fact that the dentist was the tooth puller must have been part of the initial memory representa­tion of the pair of sentences. The question was whether the elaborative inference was also part of that representation. It turned out that verifi­cation times were about the same following the first two types of pairs, but slower following the third. The conclusion, then, is that the bridg­ing inference is made immediately, in order for the subject to connect the two sentences in the text. The elaborative inference is made only when the sentence is presented for verification. A very interesting ques­tion is when and under what circumstances elaborative inferences are made in day-to-day text processing.

All inferences, whether bridging or elaborative, are based either on logic or on real-world knowledge. If a woman says that she was born in 1949, you can logically infer (in 1999) that she is 50 years old (or will be sometime before the end of the year). Most inferences, however, are based on real-world knowledge, and the trick of having successful con­versations is making sure that both speaker and hearer share enough knowledge to make appropriate inferences. Many inferences are made

on the basis of scripts (Schank & AbeJson, 1977), which are general sce­narios about common sequences of activities. For instance, most people have a "restaurant script." If someone says that Fred went to a restau­rant and ordered a steak, knowledge of the restaurant script allows the hearer to infer that Fred was served by a waiter or waitress, ordered from a menu, ate the steak, received and paid a check, and so on. Part of learning a' new job or how to function in a new institution (like a school) is acquiring scripts of how things are done there. The closer people are socially and culturally, the more shared scripts they will have. This is why people who share little information will find commu­nication more difficult than will those who share a great deal. The more information two people share, the more likely each of them will be able to judge correctly what is in the mind of the other and, therefore, what inferences that person can be relied upon to make. Later in the chapter when the role of inferences and mutual knowledge in conversations and in communication is discussed, it will be evident why shared infor­mation is so important.

In addition to creating the cohesive and coherent interpretation of texts, inferences can convey information. If someone says, "John is coming home from school this weekend. The nerd will probably spend the entire weekend at the computer," then the anaphoric resolution of John as the referent of the nerd conveys the speaker's opinion of John. Information can also be derived from inferences in sentences contain­ing because and but. Somewhere in this book there could be a sentence such as John is a good choice for the thematic role o f agent because he is ani­mate. If you did not know before you read it, you can infer from this sentence that it is, in general, good for agents to be animate. Similarly, ifI say to you, "My son-in-law is a neurosurgeon, but he's a really sweet guy," you can infer that I believe neurosurgeons are usually not sweet. Vonk and Noordman (1990) suggest that inferences of this type are not made if a person is reading material that is unfamiliar to him or her. This is unfortunate, because it could mean that students reading text­books about unfamiliar material may read at too "shallow" a level and fail to make inferences that would enhance their learning.

Much of the discussion to this point has dealt with short stretches of text, but this has been done simply to illustrate the various principles that have been introduced. In fact, people are usually involved in pro­cessing—reading or hearing—long stretches of text, whether it is a nar­rative, a lecture, or a conversation. In doing this, a hearer inconstantly

building a representation of the meaning ol the entire text, liach new sentence is integrated into that growing mental representation. The ease with which people can do this depends upon the relatedness of the individual sentences to the global structure that has been created (Hess, Foss, & Carroll, 1995). The less related an individual sentence is to the text structure that is being constructed, the more processing effort will be required to integrate it into the semantic representation of the text. Sentences that have been more difficult to integrate, or because of low relatedness are less well integrated, will be more available for recall after the text processing is complete. The same effect holds for words that are difficult to integrate with the basic meaning of a sentence (Cairns, Cowart, & Jablon, 1981). They showed, in a word probe task (like Caplan's in Chapter 7), that camera took longer to recognize if it appeared in a predictable context, such as Mary wanted to take a picture o f my baby, so she brought her camera with her, than if it appeared in an unpredictable context, such as Mary was afraid o f robbers, so she brought her camera with her. Presumably, the integration of camera is easier and more complete when it is easily connected by inference to the initial clause of the sentence than when the connection is more difficult to construct.

It is always easier for people to build a semantic representation of a text that is about something they are already familiar with. The more they know about a topic, the easier it will be to make the bridging infer­ences they need to integrate each sentence into a global representation. It is also the case that when people read (or hear) new information, they integrate it, not only with the text that is currently being processed, but also with the knowledge structures they already have about that topic. This is why advanced courses are often easier than introductory ones. In an introductory course, students are likely to know very little about the topic they are learning, and thus they do not have a knowledge base to help them integrate the texts they must process— readings and lec­tures. This problem is magnified by the fact that the person who wrote the textbooks and the person who is giving the lectures are experts in the topic, and they are formulating their texts from a background of knowledge that the student does not have. In an advanced course, on the other hand, although the course material may be more difficult, by the time students get to that level of study they have a large knowledge base to underlie their text processing.

Having Conversations

When people use language, they are involved in discourse of various kinds. Letters, stories, lectures, meetings, debates, and conversations are all types of discourse, which are, in turn, various types of texts. Although we will be concentrating on conversation, most of the prin­ciples that apply to conversation also apply to other forms of discourse. These principles, which govern the use of language, are related to prag­matics. Pragmatic principles are very different from grammatical ones. The grammar is a set of rules and principles that govern the creation of individual sentences. Pragmatic principles relate to the appropriate use of those sentences in discourse. Discourse has been characterized by philosophers of language such as J. L. Austin (1962) and John Searle(1969) as a series of speech acts. Each speech act has the following components:

Illocutionary Force. This refers to the intended function of the speech act, such as a request, a query, or transfer of information.

Locution. This refers to the linguistic form of the speech act, its structure and type, such as a question, a declarative sentence, an imperative sentence, and so forth.

Perlocutionary Force. This refers to the effect of the speech act on a listener. It can serve to change the listener's state of knowledge, cause her to respond to a request, and so on.

These distinctions are of particular importance when one analyzes nonliteral language. It is often the case that people use language when the intended meaning is very different from the basic meaning based on the words and their structural organization. One such example is the indirect request. This is when the illocutionary force of a speech act is a request, but its locution is a declarative sentence. For example, say you and your friend are in a room. You are near the window; she is not. She says, "Gee, it's getting warm in here." That is a statement (its locution), but it is intended as a request (its illocutionary force), and the effect is supposed to be for you to open the window (its perlocutionary force). Questions often have the form of a yes/no question (their location), but