Psycholinguistic Research Methods 251

See also: Deixis and Anaphora: Pragmatic Approaches; Freud S (1915). ‘Die Verdra¨ngung.’ In Gesammelte Werke Jakobson, Roman (1896–1982); Lacan, Jacques (1901– X. Frankfurt am Main: Imago. 1940–1952. 1981); Metaphor: Psychological Aspects; Metonymy; Lacan J (1955–1956). Les Psychoses. Paris: Seuil, 1981. Rhetoric, Classical. Lacan J (1957). ‘L’instance de la lettre dans l’inconscient ou la raison depuis Freud.’ In E´ crits. Paris: Seuil, 1966. Lacan J (1957–1958). Les formations de l’inconscient. Bibliography Paris: Seuil, 1998. Lacan J (1959). ‘D’une question pre´liminaire a` tout traite- Freud S (1900). ‘Die Traumdeutung.’ In Gesammelte werke ment possible de la psychose.’ In E´ crits. Paris: Seuil, II-III. Frankfurt am Main: Imago. 1940–1952. 1966. Freud S (1901). ‘Zur Psychopathologie des Alltagslebens.’ Lacan J (1964). Les quatre concepts fondamentaux de la In Gesammelte werke IV. Frankfurt am Main: Imago. psychanalyse. Paris: Seuil, 1973. 1940–1952. Lacan J (1969–1970). L’envers de la psychanalyse. Paris: Freud S (1905). ‘Der Witz und seine Beziehung zum Unbe- Seuil, 1991. wussten.’ In Gesammelte Werke, VI. Frankfurt am Main: Imago. 1940–1952.

Psycholinguistic Research Methods S Garrod, University of Glasgow, Glasgow, UK This usually involves taking the average of the values ß 2006 Elsevier Ltd. All rights reserved. of the dependent variable (response latencies) for each value of the independent variable (each list of words acquired at different ages) and establishing Psycholinguistics aims to uncover the mental repre- whether the differences associated with the different sentations and processes through which people pro- values of the independent variable are statistically duce and understand language, and it uses a wide reliable. It is now standard in psycholinguistics range of techniques to do this. The preferred psycho- experiments to test that the effects hold true both linguistic method is to carry out a controlled experi- across the range of participants used in the experi- ment. This means that the researcher manipulates an ment and across the range of linguistic materials used independent linguistic variable to control some aspect in the experiment. of language processing and then measures the effect Because the psycholinguist is interested in the of the manipulation on a dependent variable of inter- dynamics of language processing, an important dis- est. For example, consider an experiment aimed at tinction is drawn between on-line techniques, which discovering the influence of the age of acquisition of a measure variables that tap into language processing word (independent variable) on the time it takes an as it happens, and off-line techniques, which measure adult to identify that word (dependent variable). The variables related to the subsequent outcomes of pro- psycholinguist would construct two lists of words, cessing. In practice, on-line and off-line techniques one containing words learned early in life, the other compliment each other, with off-line techniques containing words learned later in life, and then mea- used to determine the outcome of interpretation and sure the effect of age of acquisition (the independent on-line techniques used to determine its time course. variable) on the time to identify the word (the depen- Another major distinction between psycholinguis- dent variable). To make sure that the independent tic techniques relates to the nature of the dependent variable is the real cause of any effect observed in variables they measure. Some experiments have be- the dependent variable, the experimenter needs to havioral dependent variables, such as those asso- carefully control that variable. For example, in the ciated with a reader’s eye movements while reading, above experiment it is important to make sure that others have neurophysiological dependent variables age of acquisition is not confounded with the length such as those associated with electrical brain activity of the word or its citation frequency, because we produced while listening to a sentence. The article know that longer words and low frequency words first considers behavioral methods and then related take longer to read. neurophysiological methods. It starts with methods Once the experiment has been run with a sufficient for the study of spoken and written language compre- number of participants and a sufficient range of lin- hension, and then methods for studying language guistic materials, the data is analyzed statistically. production and dialogue. 252 Psycholinguistic Research Methods

Behavioral Methods boosts the interpretation of the target (e.g., by speed- ing up the recognition of the target item), this is taken Common Assumptions Underlying Behavioral as evidence that there is something in common be- Methods tween the representation of target and prime. For Although there is a wide range of behavioral psycho- example, if a person is quicker to decide that giraffe linguistic methods, most depend upon the same basic is a word having just read the word tiger than having assumptions. One important assumption concerns just read the word timer, then it is assumed that this is how measurements of the time to carry out a task due to the closer semantic relationship between tiger- relate to inferences about complexity of processing. giraffe than between timer-giraffe. On the hand, Whether the timed response be an eye movement or when it hinders the interpretation of the target – this the time to answer ‘yes’ or ‘no’ to a question, it is is called negative priming as opposed to positive assumed that the complexity of the mental process is priming – this is taken to reflect some conflict be- reflected in the response latency. For example, if a tween the representations of prime and target. reader takes longer to read the fragment ‘‘... raced Priming techniques are particularly useful for estab- past the barn fell’’ when it is part of the sentence ‘‘the lishing the relationship between different linguistic horse raced past the barn fell’’ than the sentence representations used during processing and are wide- ‘‘the horse that was raced past the barn fell,’’ this is ly used in conjunction with a number of behavioral taken to reflect greater complexity in the syntactic or neurophysiological measures. analysis of the former than the latter. This assumption It is beyond the scope of this article to describe all is used when interpreting results from most on-line the behavioral techniques that have been used in psy- techniques. cholinguistics. Instead, the article concentrates A more sophisticated timing methodology investi- on techniques that have had a major impact on the gates the trade-off between speed to respond and field. accuracy of that response. This is called a speed accu- Behavioral Methods for Spoken Language racy trade-off (SAT) method. With SAT, participants Comprehension are required to make some response to a linguistic stimulus as soon as they hear a tone, presented at Cross-modal Priming A common priming technique different intervals after the presentation of the stimu- used to tap into spoken language comprehension is lus. When the interval is short, participants tend to what is called cross-modal priming. In cross-modal make many errors, and when sufficiently long, they priming the prime item is usually a word embedded in become completely accurate. So plotting response a spoken sentence or text used to prime a target item latency and accuracy across the range of tone inter- presented in written form, which is why it is called vals gives an unbiased measure of the rate at which cross-modal. Consider the problem of working out the task can be carried out. SAT techniques have been how listeners resolve ambiguous words. Cross-modal used to assess the rate at which different kinds of priming can indicate the immediate interpretation of lexical, syntactic, and semantic processing occur. an ambiguous word, such as bug, in contexts that Although it is considered a particularly refined tech- promote either one or other meaning of the word nique for establishing processing rates, it has the dis- (e.g., ‘insect’ or ‘listening device’). As participants advantage that many thousands of trials have to be listen to bug in the different contexts, they are pre- run to produce a clear SAT profile, and this will mean sented with a written word (ANT or SPY) or a non- that there will have to be many examples of each kind word (AST) and have to decide as quickly as possible of material used. So there are only a limited number whether the target is a word or not (this is called of issues that can easily be investigated with SAT. lexical decision). The question of interest is whether Another important assumption underlying many lexical decision of the targets ANTand SPY is boosted behavioral methods concerns the interpretation of by hearing the prime bug in the different contexts. priming effects. Priming techniques measure the ef- It turns out that when the visual target is presented fect of having previously processed a prime item on immediately after hearing the prime in either context the subsequent processing of a target item. The prime it promotes lexical decision for both targets. So this might be a word with a particular form or meaning indicates that both interpretations (‘insect’ and ‘lis- or it might even be a whole sentence with a particular tening device’) are immediately activated irrespective syntactic structure. The rationale behind priming of the disambiguating context. However, if the visual techniques is that any influence of prime on the target is presented at a slightly later point (about subsequent processing of the target must reflect 200 ms after the prime), only the target related to some relationship between the mental representations the contextually appropriate interpretation (i.e., of prime and target items. Typically, when a prime ANT for ‘insect’ or SPY for ‘listening device’)is Psycholinguistic Research Methods 253 primed. Cross-modal priming therefore indicates that in the box when there is either one or two frogs in all meanings of ambiguous words are accessed imme- the scene. Analyses of eye movements demonstrate diately on encounter, but then only the contextually that viewers look at the frog which is on a napkin appropriate meanings are retained as comprehension more if there is also a frog that is not on a napkin than proceeds. if there is no other frog present. From this it can be Cross-modal priming has been used to investigate inferred that the visual context drives syntactic dis- many aspects of spoken language comprehension. ambiguation (i.e., it supports the reading in which on These include influences of prior discourse context the napkin syntactically modifies frog rather than on lexical processing, resolution of anaphoric refer- being a syntactic argument of put) at an early stage ences (e.g., pronouns), and morphological analysis. in processing. The technique has the advantage that it can tap The visual world paradigm has been used to inves- into spoken language comprehension as it occurs tigate a wide range of issues in spoken language com- and often uncovers aspects of processing of which prehension. These include lexical access, resolution of the subject is completely unaware, such as in the anaphoric pronouns, development of strategies for ambiguity experiment described above. semantic and syntactic processing, language develop- ment, and even language processing in dialogue. It The Gating Technique Another technique for estab- offers a precise indication of when listeners integrate lishing when listeners interpret words in relation information from a linguistic utterance with that in to information in the speech stream is gating. The the visual world and tends to show that comprehen- gating procedure involves presenting increasingly sion is both incremental and immediate in relation to long fragments of speech and measuring when listen- most levels of linguistic analysis. ers can interpret the speech appropriately. For exam- ple, if you want to know at what point a listener can Behavioral Methods for Written Language accurately recognize the word cathedral, you can Comprehension present them with the spoken fragments /ca/, /cath/, There are a variety of techniques that tap into the /cathe/ , /cathed/ ... and record the listener’s judge- time course of written language comprehension. ments for each fragment. The shortest fragment that can be correctly identified defines the point at Self-paced Reading One class of techniques is self- which there is sufficient information in the speech paced reading. The reader determines the rate at for identification. This technique can be used to esti- which written material is presented and the experi- mate the earliest point at which the listener could iden- menter records the rate of presentation. A reader tify a word and so make it possible to test whether this might be required to pace himself or herself sentence predicts the recognition time for that spoken word. by sentence, phrase by phrase, or word by word. For example, in the word-by-word procedure, a word is The Visual World Paradigm One of the most effec- presented, and as soon as the reader has understood tive on-line measures is eyetracking (recording the it, he or she presses a key to trigger presentation of the precise pattern of eye fixations during comprehen- next word. The sequence is then repeated until all the sion). Because visual attention is strongly controlled text has been read and the time to read each word is by where a person is currently looking, eyetracking recorded. This kind of technique has been used to can be used to indicate what a person is attending to study syntactic analysis, discourse comprehension at any point during comprehension and for how long processes and in particular resolution of anaphors. It they attend to it. The technique can either be used for gives a good indication of when a reader encounters reading research (see below), in which case the focus difficulty in comprehension, but is limited according of attention corresponds to the words being looked at to the size of linguistic unit being presented. Whereas any time or it can be used to measure which part of a larger units such as whole sentences can be read at a scene a participant attends to as they interpret spoken normal rate during self-paced reading, smaller units utterances about that scene. The latter technique is like words tend to be read much more slowly in self- usually referred to as the visual world paradigm. paced reading tasks. Hence, when the technique has A classic example of the use of this technique is in high on-line resolution (e.g., when word by word), it determining how listeners deal with syntactic ambi- also interferes most with the normal reading process. guities that arise during comprehension. For example, This is not a problem with the eyetracking technique participants listen to instructions that are initially described below. consistent with two syntactic analyses while they view a scene containing a small number of objects. Rapid Serial Visual Presentation (RSVP) A slightly They might be asked to Put the frog on the napkin different technique for presenting written language 254 Psycholinguistic Research Methods uses what is called rapid serial visual presentation Moving window studies indicate that readers only (RSVP). With RSVP, readers see sequences of words take in information from a limited region of text at in the center of a computer screen presented at a fixed any time during reading. This means that any extra fast rate. The experimenter then has the reader carry time spent fixating the region must reflect processing out an additional task, such as identifying a word in difficulty associated with that region of text or previ- the sequence or trying to recall the sequence of words, ously fixated regions of text not completely processed which indicates how comprehension is limited by the but still held in memory. rate of presentation. This technique has been used Eyetracking has been used to study a wide range of to investigate lexical, semantic, and syntactic proces- linguistic processes, including lexical access, resolving sing. Like word-by-word self-paced reading, RSVP lexical ambiguities, syntactic analysis, and various may well interfere with normal language processing. discourse processing phenomena, such as anaphora resolution. It is particularly effective in determining Eyetracking During Reading The least interfering precisely when the reader makes a decision about on-line behavioral technique for written language some aspect of the linguistic input during sentence comprehension is eyetracking. During reading, the or discourse processing. For example, when presented eye moves in a systematic way. There are brief fixa- with the sentence We like the city that the author tions in which gaze stays on the same letter inter- wrote unceasingly and with great dedication about spersed with fast movements called saccades during readers spend longer fixating the verb wrote as com- which the gaze moves to another letter or word of the pared with the same verb in the fragment We like the text. For a skilled reader, 9 out of 10 saccades move book that the author wrote unceasingly and with the gaze from left to right to sample new material great dedication about. This shows that readers im- from the text, whereas 1 out of 10 saccades return the mediately attempt to integrate each word of the sen- point of gaze to previously read material (these are tence with the prior discourse and hence they detect called regressions). The duration of fixations and the the temporary anomaly produced by wrote in the first length and direction of saccades (i.e., forward or but not in the second sentence. backward movement of the gaze) directly reflect the Eyetracking is a particularly effective technique ease or difficulty of the reading process. Furthermore, because it does not interfere with the normal process they indicate the precise word in the text that is of reading. A similar claim is made for some neuro- causing reading difficulty because attention is only physiological techniques such as ERP, described later. given to the word currently fixated. The limited span of attention during reading can be Behavioral Methods for Spoken Language demonstrated using the moving window technique in Production which a computer program controls dynamically the window of text presented to the reader as a function Until the last decade, language production was not of where they are fixating. For example, with an a central topic in psycholinguistics. This was partly asymmetric 12-letter window, the 4 letters to left because researchers did not have on-line methods for and the 8 letters to the right of where the reader is studying production in properly controlled experi- fixating will be displayed as normal, whereas all the ments. This meant that the pioneering research on remaining text will be converted into random letters. language production depended on off-line techni- The window of text together with its surround of ques, such as the study of speech errors. However, random letters then changes as the point of fixation more recent work in language production has been changes. One can reduce the size and form of the text influenced by on-line techniques. We consider both window and measure when it begins to affect reading kinds of technique below. rate. It turns out that normal reading is quite possible when the window only contains the word currently Analysis of Speech Errors Speech error data has fixated plus the first three letters of the next word on been used to draw many interesting conclusions the line. However, there is a proviso that the material about the nature of language production. For exam- around the window must retain the spaces between ple, it was observed that substitution errors (e.g., the words in the original text. When the window saying if I was done to that rather than If that was arrangement is reversed so that the window contains done to me) tend to always involve the same syntactic random letters and the surround contains the normal classes (e.g, the pronouns me and that). Also, as the text, readers encounter difficulty. With a reverse win- example shows, the lexical substitution does not al- dow of only 11 letters in width, reading becomes ways involve a syntactic substitution. Otherwise, the almost impossible. error would have produced If me was done to that. Psycholinguistic Research Methods 255

Such findings provide evidence that in production, captain-major). The question is whether participants words are chosen before they are strung together can use the implicit prime of the shared first syllable to make up an utterance and that this occurs before to speed up articulation. It turns out that they can, the words are marked for syntactic case. and more interestingly, the implicit priming only Speech errors have been used to argue for an over- works for the first syllable in the word. When given all organization of speech production into separate a comparable list in which the second syllable of message planning, lexical and grammatical assembly, the second word is shared (e.g., single-murder, and phonological processing components. There have place-ponder, fruit-boulder), there is no articulatory also been some attempts to develop techniques to benefit. elicit speech errors experimentally by priming the There are also techniques for studying priming at errors in a similar fashion to that used in tongue the syntactic level in production which use a variant twisters. Many of the conclusions drawn from the of the picture naming procedure. A typical study analysis of speech errors have been supported by might involve participants describing a sequence of results from more recent on-line techniques such as pictured events using a verb indicated at the bottom picture naming. of each picture. Interleaved between these descrip- tions, the participant checks descriptions they hear Picture Naming A particularly influential on-line against another series of pictured events. By using technique for studying language production is to ditransitive verbs such as give, which take either measure the time it takes participants to name pic- prepositional objects (give the picture to Mary) or tures of objects or events. Picture naming is common- double objects (give Mary the picture), it is possible ly combined with some form of priming task (see next to study syntactic priming independent of semantic section). It can also be combined with other techni- priming. The question is whether having just heard ques such as eyetracking to give a more precise indi- The sailor gave the banana to the nun participants cation of how words are accessed during language are more likely to describe their next picture as production. For example, a researcher might want The clown handed the book to the pirate than The to know whether speakers wait until they have all clown handed the pirate the book? It turns out that the words available before they start to articulate they are. the first word in an utterance. To find out, the re- Behavioral Methods for the Study of Dialogue searcher could have participants name two pictured objects A and B in a phrase of the form A and B. If the As with language production, it is only quite recently time to start articulating the name of A is unaffected that psycholinguists have begun the experimental in- by the time to access the name for B as indicated by vestigation of language processing during dialogue. the time to articulate B in isolation, then the research- Again there is what has sometimes been called the er can conclude that the speech articulation process problem of exuberant responses. Because dialogue is begins before both word forms have been accessed. inherently spontaneous, how can the experimenter In fact, the evidence supports this conclusion. exert the control required for a sound experiment? One way around this problem is to set up a task Priming Techniques in Language Production An that controls what interlocutors can talk about. One important issue in language production concerns the such task is the referential communication task. One extent to which utterances are formulated incremen- participant is required to describe a series of arbitrary tally one unit at a time according to different levels of visual patterns such that their partner who is able representation (semantic, syntactic, and phonologi- to reply can identify each pattern from the set that cal). Priming techniques have been used to address they have been given. This technique has been used to this and other related questions. For example, at the investigate how feedback from the listener affects phonological level words could be assembled for the nature of subsequent references made by an articulation either as complete packages or incremen- interlocutor. More complex dialogue tasks have had tally as sequences of distinct phonological units. A so- conversational partners describe routes on a map or called implicit priming technique has been developed positions in a maze as part of some other cooperative to test this. The participant has to learn sets of pairs activity. More recently, researchers have started to of words, such that when given the first word in the record interlocutors patterns of eye movements pair he or she names the second word as quickly as while carrying out some version of the referential possible. Crucially, in one condition the second words communication task. In this way, it is possible to always share the same first syllable (e.g., single-loner, combine aspects of the visual world paradigm with place-local, fruit-lotus), whereas in the other condi- those of the interactive referential communication tion they do not (e.g., single-loner, signal-beacon, task. 256 Psycholinguistic Research Methods

Neurophysiological Methods – ERP, word was presented. This process is then repeated fMRI, and MEG over a number of trials. Because the data from each trial will contain irrelevant electrical activity, the ex- Techniques for measuring the neurophysiological cor- perimenter takes the average potential across the set relates of language processing have recently increased of related trials. In this way, the irrelevant ‘noise’ the psycholinguist’s methodological armory. Three par- information can be filtered out. What remains is an ticular measurement techniques have been used. The ERP waveform with identifiable peaks and troughs of first measures electrical activity at the scalp using elec- voltage. These peaks and troughs are taken to reflect tro-encephalography (EEG) to produce what are called the activity of bundles of nerve fibers in particular event-related brain potentials (ERPs). The second parts of the brain. measures changes in brain blood flow associated with In practice, ERP researchers try to identify the neural activity using functional magnetic resonance characteristic peaks and troughs and establish how imaging techniques (fMRI), and the third measures they might relate to concurrent processing. One well- changes in magnetic fields associated with the electrical established peak is called the N400, which corre- activity in the brain using magneto-encephalography sponds to a negative component of the wave occurring (MEG). Each technique has its own advantages and approximately 400 ms. after presentation of the word disadvantages as a psycholinguistic tool. that triggers it (it is a peak because ERP researchers fMRI signals give precise information about the conventionally plot negative values upward and posi- area in the brain associated with the particular activi- tive values downward). The N400 has been associated ty, so fMRI has proved useful for neurolinguistic with processes of conceptual or semantic integration investigation. The disadvantage with the technique of words into their sentential contexts. For example, for psycholinguistics is that it is not good for estab- when you measure ERPs elicited by the words eat, lishing the time course of the neural activity. This is drink, or cry in the context ‘‘The pizza was too hot because it takes time for the neural activity to produce to eat/drink/cry,’’ then the N400 is larger for cry than changes in the blood flow. By contrast, ERPs provide drink and larger for drink than the contextually ap- precise information about the time course of the neu- propriate eat. This pattern is consistent with the ral activity, but it is difficult to establish the source of idea that the N400 reflects conceptual integration the activity. This is because the ERP signal is affected (with eat being integrated into the sentence better by all sorts of irrelevant factors such as the thickness than drink) as well as semantic integration (drink is of the skull or interactions between signals from dif- semantically related to eat whereas cry is not). ferent areas of the brain. Finally, the most recently Another characteristic trough is called the P600 (a developed technique, MEG, offers good localization positive change occurring about 600 ms after the with similar temporal resolution to ERP. Neverthe- triggering word). Unlike the N400, the P600 has less, like ERP signals, MEG signals are only sensitive been associated with syntactic integration processes. to activity in neural structures with particular orien- For example, the word was when presented in tations, and neither technique is easy to apply when the ungrammatical sentence The doctor forced the there is contemporaneous motor activity, such as eye patient was lying produces a much larger P600 than movements or articulation during speech. was in the sentence The doctor thought the patient The most influential psycholinguistic research to was lying. Because these two wave forms, N400 and date has tended to use ERP because it is both relative- P600, seem to reflect different kinds of processing, ly cheap to run ERP experiments and the technique ERP can be used to establish the precise time course of offers similar temporal resolution to that of behav- these different processes. So ERP complements other ioral measures such as eye tracking. This article con- on-line measures such as eyetracking, which do not centrates mainly on the use of ERPs, but it will also differentiate between different kinds of psycholin- say something about the use of MEG. guistic processes in this way. ERPs have been used to address a wide range of Event-related Brain Potentials (ERPs) questions both about early stages of lexical proces- ERPs are derived from measurements of small changes sing and more general syntactic and semantic process- in voltage at different points across the scalp. They es in language comprehension. It is best suited for the are called event-related potentials because they are study of processes that immediately follow presenta- analyzed in relation to the onset of a triggering tion of the triggering stimulus because the ERP signal event. For instance, to derive an ERP that reflects becomes increasingly noisy over time. This means word identification processes, the experimenter pre- that it is not such a good technique for studying sents a word on a computer screen and measures the such things as syntactic re-analysis or integration of changes in scalp voltage from the point at which the a sentence into the discourse context. Psycholinguistics: History 257

Magneto-encephalography (MEG) they tap into language processing as it happens. Behavioral measures, such as eyetracking, and neuro- MEG has only recently begun to be used in psycho- physiological measures, such as ERP, are particu- linguistics. MEG has some of the advantages of both larly effective for measuring the time course of fMRI techniques and EEG techniques because it language comprehension. For language production enables precise source localization like fMRI and studies, picture naming and priming techniques have has fine temporal resolution like ERP. It also comple- been especially effective. ments ERP. Whereas ERP electrical signals can only be picked up from nerve bundles that are in particular orientations with respect to the surface of the brain, See also: Dialogue and Interaction; Evoked Potentials; MEG magnetic field signals can be picked up from fMRI Studies of Language; Magnetoencephalography; nerve bundles that are orthogonal to those giving ERP Psycholinguistics: Overview; Reading Processes in signals. For these reasons, many researchers are par- Adults; Speech Errors: Psycholinguistic Approach; Spo- ticularly optimistic about MEG as a psycholinguistic ken Language Production: Psycholinguistic Approach. method used together with ERP. One particularly interesting application has been using MEG to establish the relationship between neu- Bibliography ral representation and linguistic form. The technique depends upon what has been called mismatch nega- Carreiras M & Clifton C E (eds.) (2004). The on-line study tivity. It was observed that as the same items are of sentence comprehension: Eyetracking, ERP and repeatedly presented to subjects, so the MEG signa- beyond. Hove: Psychology Press. Garrod S (1999). ‘The challenge of dialogue for theories of ture associated with their processing is automatically language processing.’ In Garrod S & Pickering M J (eds.) reduced (probably because of neuronal habituation). Language processing. Hove: Psychology Press. 389–415. However, when a new item is presented, the signal Levelt W J M (1989). Speaking: from intention to articula- returns to normal. This happens irrespective of any tion. Cambridge, MA: MIT Press. behavior on the part of the subject. Mismatch nega- Osterhout L & Holcomb P J (1995). ‘Event-related poten- tivity can therefore be used to establish the degree to tials and language comprehension.’ In Rugg M D & which different items are processed in the same way. Coles M G H (eds.) Electrophysiology of mind: event- The greater the resumption of the activity (i.e., mis- related brain potentials and cognition. Oxford: Oxford match negativity), the more different the neurological University Press. 171–216. processing of the new item. In this way, mismatch Pulvermu¨ ller F (2001). ‘Brain reflections of words and their negativity can be used in a similar fashion to priming meanings.’ Trends in Cognitive Sciences 5, 517–524. Rayner K & Sereno S C (1994). ‘Eye movements in reading: techniques to explore the neurological representation Psycholinguistic studies.’ In Gernsbacher M A (ed.) of different aspects of a linguistic stimulus. Handbook of Psycholinguistics. San Diego: Academic Press. 57–82. Summary and Conclusion Trueswell J C & Tanenhaus (eds.) (2004). Approaches to studying world-situated language use: bridging the Psycholinguistic techniques differ according to the language-as-product and language-as-action traditions. kind of variables measured and the extent to which Cambridge, MA: MIT Press.

Psycholinguistics: History G Altmann, University of York, York, UK literature. This proliferation was marked by the ß 2006 Elsevier Ltd. All rights reserved. founding in 1962 of the Journal of Verbal Learning and Verbal Behavior (which subsequently, in 1985, became the Journal of Memory and Language). Why Our faculty for language has intrigued scholars for the original journal was so titled, and why its title centuries. Yet most textbooks assume that psycho- presents us with a historical paradox, will become linguistics has its origins in the late 1950s and 1960s, clearer as this review unfolds. The review’s purpose and that nothing of note contributed to its evolution is to consider how the present-day state of the art before then. In some respects this is true, in that it was evolved. In so doing, it will touch briefly on ancient only then that psycholinguistics began to proliferate Greek philosophy, 19th century neuroscience, 20th as an identifiable discipline within the psychology century psycholinguistics, and beyond. It will 258 Psycholinguistics: History consider approaches to the brain as practiced in both The psycholinguistic endeavor is to uncover the ancient Egypt and modern neuroscience. It will be mental processes that are implicated in the acquisi- necessarily selective, in order to make some sense tion, production, and comprehension of language. of the historical developments that contributed to Just as psychology is the study of the control of psycholinguistic science. behavior, so psycholinguistics is the study of the con- trol of linguistic behavior. A part of any psycholin- guistic theory of mental process is an account of what From the Ancient Egyptians to the constitutes the input to the mental process – that is, Greek Philosophers what information is operated upon by those process- es. While Plato was of course correct that the form of The earliest to write about language and the brain the real-world object dictates the form of the sensory were the ancient Egyptians – the first to write about image presented to the allegorical prisoner, the mental anything at all. A catalog of the effects of head injury processes involved in that prisoner’s use of language (and injuries lower down the body also) exists in what can operate only on mental derivatives of that sensory is now referred to as the Edwin Smith Surgical Papy- image. There may be properties of the real-world rus, written about 1700 B.C. The writer (believed to object (such as color, texture, and density) that are have collected together information spanning per- not represented in their shadow-forms, and thus men- haps another 1000 years before) referred there to tal processes that might otherwise (outside the cave) what is presumed to be the first recorded case of develop sensitivity to those properties need never de- aphasia – language breakdown following brain trau- velop such sensitivities if constrained to living a life ma. However, the Egyptians did not accord much inside the cave. But while the shadows would not significance to the brain, which unlike the other permit the distinction between, say, a tennis ball and organs of the body, was discarded during mummifi- an orange, the contexts in which the shadows were cation (it was scraped out through the nose). They experienced, or their names heard, would distinguish believed instead that the heart was the seat of the soul between the two – mental sensitivities would develop, and the repository for memory, a view largely shared but they would not necessarily be grounded in the by the Greek philosopher Aristotle (384–322 B.C.) – perceptual domain. These distinctions, between the a somewhat surprising position to take given that actual world and our experience of the world, and he was a student at Plato’s Academy and that Plato between an object or word and the context in which (427–347 B.C.) believed the brain to be the seat of that object or word might occur, led other philoso- intelligence. phers, most notably Wittgenstein in his Philosophical Plato was possibly the earliest to write at length on Investigations, to propose that the meaning of a word language (where others may have spoken, but not is knowledge of its use in the language – that is, written). Certainly, his writings were the most influ- knowledge of the contexts in which it would be ap- ential with respect to the philosophy of language propriate to utter that word, where such knowledge is and the question ‘what does a word mean?’ Plato, in shaped by experience. We return to this theme when his Republic, considered the meaning of words in we consider in more detail the more recent history his Allegory of the Cave (as well as in Cratylus). In this (and possible future) of psycholinguistics. allegory, a group of prisoners have been chained all their lives within a cave. All they see are the shadows The Earliest Empirical Studies of objects cast upon a wall by the flames of a fire. They experience only those shadows (in much the The pre-history of psycholinguistics (up until the same way that we can only experience the results of 19th century) was dominated by philosophical con- our sensory percepts), and their language similarly jecture. The term dominated is used loosely here, as describes only those shadows. Plato noted that when there was no systematic and ongoing questioning of using a word, the prisoners would take it to refer to the relationship between mind and language, or in- the shadows before them, when in fact (according to deed, brain and language – there was no community Plato), they would refer not to objects in the shadow of researchers asking the questions. But modern-day world, but (unbeknown to the prisoners) to objects in psycholinguistics is dominated not by philosophy (al- the real world. Thus, for Plato (and a host of more though it had its moments), but by experimental in- contemporary philosophers, from Frege to Puttnam), vestigations that measure reaction times, monitor eye the true meaning of a word – its reference – is external movements, record babies’ babbles, and so on. Its pre- to the person who, by using the word, is attributing history lacks such experimentation. This is not to say meaning to it. But why should it matter what a word that no experiments were performed. Certainly, there refers to? were isolated cases, generally of a kind that would not Psycholinguistics: History 259 be tolerated in the modern age. Indeed, one of the accompanied by a frontal lesion, he would give up his most widely replicated studies (if one is to believe the (and his intellectual forbearers’) belief in the localiza- historians) is a study that was carried out on at least tion of language. In the audience was Paul Broca, three and possibly four independent occasions be- after whom are named Broca’s aphasia and, within tween the 7th Century B.C. and the 16th Century A.D. the left frontal lobe, Broca’s area. Broca was struck by In each case, some number of babies were apparently Aubertin’s empirical challenge, but at the same time brought up in isolation (except for carers who were realized that craniology (Gall’s lasting influence on either mute, or instructed not to speak), with the aim his students) could not provide the proof that was of the experiment being to discover what language, if required to establish a link between language loss any, the children would grow up speaking. The results and cerebral localization – only anatomical inspec- varied. The Egyptian Pharaoh Psamtik (7th C.B.C.) was tion of the brain could do that. Coincidentally, within credited by Herodotus with discovering that they a few days he was presented with a patient suffering spoke Phrygian. The Roman emperor and German from speech loss who died a few days after that. king Frederick II (1194–1250 A.D.) carried out a simi- Broca’s postmortem analysis of this patient’s brain lar study, but all the infants died. King James IV (and the damage to what is now referred to as Broca’s (1473–1513 A.D.) is supposed to have performed a area), coupled with earlier observations made by Marc similar experiment on the island of InchKeith, al- Dax (on right hemiplegia and its correlation with though it is likely that this study never in fact took speech loss), but published at the same time, established place (the fact that the children are reported to have the anatomical validity of the localization hypothesis. emerged from their isolation speaking Hebrew is one About 10 years later (in 1874), Carl Wernicke pub- reason to doubt the truth of the story). And finally, lished his work on ‘sensory aphasia’ (deficits in the Akbar the Great (1542–1605), the grandfather of comprehension of language). This work was consider- Shah Jahan who built the Taj Mahal, similarly failed ably enhanced by Wernicke’s student Ludwig to discover man’s ‘natural language’ (although there Lichtheim who, in 1885, produced a schematic (cf. is some suggestion that in this case, the infants ac- a ‘model’) of how three interlinked centers in the quired a form of signed language inherited, in part, brain are implicated in aphasia: Broca’s(the‘center from the infants’ carers). of auditory images’), Wernicke’s(the‘center of motor images’), and a diffusely located ‘concept center.’ The 19th Century Emergence of the Lesions to each of these areas, or to the connections Cognitive Neuropsychology of Language between them, produce different kinds of aphasias. Most interesting of all, his schematic enabled him to The first systematic studies of the relationship between predict disorders that had not yet been described. This language and brain were conducted in the 19th cen- ability of a conceptual ‘model’ to make as yet untested tury. This is probably the earliest point in the history predictions is a theme we shall return to. of psycholinguistics from when a progression of stud- ies can be traced, with one author building a case on The Early 20th Century Influence of the basis of earlier studies coupled with newer data. Behaviorism The protagonists at this time were Gall, Boulliard, Aubertin, Broca, Wernicke, and Lichtheim, to name a By the end of the 19th century, the study of language few. None of them would be described as ‘psycholin- began to change, as did the study of psychology guists,’ but to the extent that their work (like modern- more generally. Interest in the psychology (as op- day cognitive neuroscientists) informed accounts of posed to philosophy) of language shifted from being the relationship between brain and language, they are primarily (or even solely) concerned with its break- no less a part of the history of psycholinguistics than down to being concerned also with its normal use. are the linguists, philosophers, psychologists, and cog- Wilhelm Wundt in Die Sprache (published in 1900) nitive scientists who have influenced the field through stressed the importance of mental states and the their own, sometimes radically different, perspectives. relationship between utterances and those internal Franz Gall is perhaps better known for his work on states. William James similarly (at least early on) phrenology, but he believed that language function saw the advantages of introducing mental states into was localized in the anterior parts of the brain. His theories of language use (see his 1890 Principles of student Jean Boulliard collected clinical evidence in Psychology, in which several contemporary issues in support of Gall’s theory, and in turn, Boulliard’s stu- psycholinguistics are foreshadowed). But the early dent Ernest Aubertin did the same. It was at a meeting 20th century was a turbulent time for psycholinguis- in April of 1861 that Aubertin made his beliefs plain: tics (as it was for psychology): J. B. Watson argued If a case of speech loss could be found that was not that psychology should be concerned with behavior 260 Psycholinguistics: History and behavioral observation, rather than with cons- the infinite productivity or systematicity of language. ciousness and introspection (the Wundtian approach). With Chomsky, out went Bloomfield, and in came And whereas Wundt had argued that a psychology mental structures, ripe for theoretical and empirical of language was as much about the mind as it was investigation. Chomsky reintroduced the mind, and about language, behaviorists such as J. R. Kantor specifically mental representation, into theories of argued against the idea that language use implicated language (although his beliefs did not amount to a distinct mental states. For Kantor, the German men- theory of psychological process, but to an account of talist tradition started by Wundt was simply wrong. linguistic structure). So whereas Skinner ostensibly Even William James turned away from Wundtian eschewed mental representations, Chomsky appar- psychology. Thus, the behaviorist tradition took hold. ently proved that language was founded on precisely The late 19th and early 20th centuries were a time such representation. Some later commentators took of great change in linguistics, too. The 19th century the view that the Chomskyan revolution threw out the had seen the emergence of the Neogrammarians, a associationist baby with the behaviorist bathwater. group that studied language change. They were inter- Behaviorism was founded on associationism. Behav- ested in how the sounds of different languages were iorism was ‘out,’ and with it, associationism. Symbol- related, and how within a language, the sounds ic computation was ‘in,’ but with it, uncertainty changed over time. They were less interested in what over how the symbolic system was acquired. It was a language ‘looked like’ at a particular moment in not until the mid-1980s that a new kind of revolution time. This changed at the beginning of the 20th took place, in which the associationist baby, now century when Ferdinand de Saussure brought struc- grown up, was brought back into the fold. The inter- ture into the study of language. He introduced the vening 20 years were typical teenage years – full of idea that every element of language could be under- energy, punctuated by occasional false hopes that stood through its relation to the other elements (he nonetheless proved essential to the maturation introduced syntactic distinctions that are still cen- process. tral to contemporary linguistics). In the 1930s, the Two years before his review of Verbal Beha- Bloomfieldian school of linguistics was born, with the vior, Chomsky had published Syntactic Structures,a publication in 1933 of Leonard Bloomfield’s Lan- monograph devoted to exploring the notion of ab- guage. Bloomfield reduced the study of language stract grammatical rules as the basis for generating structures to a laborious set of taxonomic procedures, sentential structure. According to Blumenthal in his starting with the smallest element of language – the 1970 account of the history of psycholinguistics, phoneme. In doing so, Bloomfield firmly aligned Chomsky’s departure from the Bloomfieldian school the linguistics of the day with behaviorism. And was too radical for an American publisher to want to just as behaviorism eschewed mental states in its publish a lengthy volume that Chomsky had written study of psychology, so the Bloomfieldian tradition outlining the new approach, and only Mouton, a eschewed psychology in its study of language. The European publisher (and presumably more sympa- study of language was firmly caught between the pro- thetic to the tradition that Chomsky was advocating) verbial rock and a hard place – between behaviorism would publish a shorter monograph based on an on the one hand and taxonomy on the other. Mental undergraduate lecture series he taught at MIT. In states were, the argument went, irrelevant – whether fact, this is not quite accurate (N. Chomsky, personal with respect to psychological or linguistic inquiry. communication); Chomsky had indeed written a The behaviorist tradition culminated (with respect longer volume (subsequently published in 1975), to language) with B. F.Skinner’s publication in 1957 of and it is true that initial reactions to the manuscript Verbal Behavior. Here, Skinner sought to apply behav- were negative (but, according to Chomsky, not unrea- iorist principles to verbal learning and verbal behavior, sonable), but Syntactic Structures was not a compro- attempting to explain them in terms of conditioning mise brought about through Chomsky’s search for a theory. Verbal behavior (and Verbal Behavior) proved publisher; he had not, in fact, intended to publish it. to be the final battleground on which the classical Instead, Cornelis van Schooneveld, a Dutch linguist behavorists and the mentalists would clash. and acquaintance of Chomsky’s who was visiting MIT and happened to edit a series for Mouton, sug- The Mid-20th Century and the gested that Chomsky write up his class notes and Chomskyan Influence publish them. This he did, and modern linguistics was born. Psycholinguistics became caught up, al- In 1959, Chomsky published a review of Skinner’s most immediately, in its wake. Verbal Behavior. He argued that no amount of condi- Chomsky’s influence on psycholinguistics cannot tioned stimulus-response associations could explain be overstated. He drew an important distinction Psycholinguistics: History 261 between ‘competence,’ or the knowledge we have Thus, despite the grammaticality of ‘the horse raced about a language, and ‘performance,’ the use of that past the barn fell’ (cf. ‘the car driven past the garage language (a distinction that was reminiscent of Saus- crashed’), the preference to interpret ‘raced’ as a main sure’s earlier distinction between langue and parole). verb (instead of as a past participle equivalent to Both, he claimed, arise through the workings of the ‘driven’) is so overwhelming that the sentence is human mind – a mind, which furthermore is innately perceived as ungrammatical (and the preference is enabled to learn the structures of human language then said to induce a ‘garden path’ effect). Evidently, (although not everyone agreed with the arguments grammaticality and processability are distinct mental for a language acquisition device akin to a mental phenomena. organ – a concise summary of the counterarguments was written by Bates and Goodman (1999)). It is On the Influence of the Digital Computer perhaps surprising that against the backdrop of Syntactic Structures and Chomsky’s Review of The 1970s saw enormous growth in psycholinguis- Skinner’s Verbal Behavior, a new and influential tics. Advances were made across a wide range of journal dedicated to research into the psychology of phenomena, including the identification of both language should nonetheless, in 1962, give itself a printed and spoken words, the reading process, sen- title (the Journal of Verbal Learning and Verbal tence comprehension (with much of the emphasis on Behavior) that firmly placed it in the behaviorist the resolution of ambiguities of the ‘garden path’ tradition. kind), and the mental representation of texts. Wheth- er there was a ‘spurt’ in the number of publications From Linguistic Competence to is contentious, because although there undeniably was such a spurt, the whole of psychology experienced Psychological Performance the same rapid growth. It would be wrong, however, Chomsky’s theories of grammar were theories of to attribute all this advancement to the influence of competence, not performance. And yet, his work on Chomsky. The demise of behaviorism played a part transformational grammar initiated a considerable (and certainly Chomsky played a part in that demise), research effort in the early 1960s to validate the but so did the advent in the 1950s of the digital psychological status of syntactic processing (the con- computer. The ‘mind-as-computer’ metaphor had a struction of representations encoding the dependen- subtle but pervasive influence on both psycholinguis- cies between the constituents of a sentence). Many tics and the study of cognition generally. Computer of these studies attempted to show that perceptual programs worked by breaking down complex beha- complexity, as measured using a variety of different viors into sequences of simpler, more manageable tasks, was related to linguistic complexity (the so- (and hence more understandable) behaviors. They called Derivational Theory of Complexity). However, relied on symbol manipulation and the control of whereas the syntactic structures postulated by trans- information flow. They distinguished between differ- formational grammar did have some psychological ent levels of explanatory abstraction (the high-level reality, the devices postulated for building those struc- programming language, the assembly code, and the tures (e.g., the transformations that formed a part of flow of electrical currents around the hardware). And the grammatical formalism) did not. It soon became perhaps most important of all to the empirical psy- apparent that the distinction between competence chologist, they enabled novel predictions to be made and performance was far more important than origi- that might not otherwise have been foreseen had the nally realized – the linguists’ rules, which formed a ‘model’ not been implemented in full; complex inter- theory of competence, did not make a theory of psy- actions among the components of a program were not chological process. easy to foresee. Subsequently, the emphasis shifted toward exami- The influences of the digital computing revolution nation of the psychological, not linguistic, mechan- were felt in different ways. Some were direct, with isms by which syntactic dependencies are determined researchers building computer simulations of mental (a process referred to as parsing). In a seminal paper behavior (in the growing field of Artificial Intelligence, published in 1970, Thomas Bever pointed out that in several language ‘understanding’ programs were writ- cases of ambiguity, when more than one structure ten, some of which are still relevant 35 years later – e.g., (i.e., dependency relation) might be permissible, there Terry Winograd’s SHRDLU program written in appear to be consistent preferences for one interpre- 1968–1970). Other influences were indirect, coming tation rather than another. This consistency appeared to psycholinguistics via philosophy. One such exam- to hold not only across different examples of the same ple was Jerry Fodor’s Modularity of Mind hypoth- kind of ambiguity, but across different people, too. esis (from 1983). One simplified interpretation of 262 Psycholinguistics: History this hypothesis (it was interpreted in different ways by Importantly, and in contrast to the ideals of the different researchers) was that there are two alterna- behaviorist traditions, neural networks can develop tive ways of theorizing about the mind: one is to internal representations. assume it is incredibly complex and that multiple Several connectionist models had profound effects sources of information interact in multiple ways, and on developments in psycholinguistics. TRACE, for the other is to assume that it can be broken down into example, developed by McClelland and Jeff Elman a number of modules, each of which performs some in the 1980s, was a model of spoken word recogni- particular function and is ‘blind’ to the workings of tion that formed the focus of empirical research for the other modules (perhaps taking as input the output a good 20 years after its inception. But TRACE did of one or more of those other modules). Fodor argued not learn anything – it was hardwired. An extremely that certain aspects of cognition were modular (the influential model that did learn by itself was de- input systems), and certain others were not (central scribed by Elman (1990), who showed how a partic- processes). This hypothesis had considerable influence ular kind of network could learn the dependencies in psycholinguistics, and for a time (the mid-1980s that constrain the sequential ordering of elements to early 1990s), hypotheses were evaluated accord- (e.g., phonemes or words) through time. In effect, it ing to whether they were modular or not. There seemed learned which kinds of word could follow which little agreement, however, on where one drew the other kinds of word (hence, it was a statistical model, boundaries (for example, was spoken language recog- because it encoded the statistics of the language it nition a part of an input system? If it was, how could was trained upon). Interestingly, it developed internal ‘higher-level’ knowledge of the context in which representations that appeared to resemble gram- the language was being interpreted influence the mod- matical knowledge; words that occurred in similar ular and encapsulated recognition process? – Some sentential contexts came to evoke similar internal argued it could not, while others argued it could). representations (that is, internal patterns of activity It was about this time, in seeming opposition to when the word was presented to the network) – and the trend toward symbolic computation, that a new because words of the same grammatical category tend computationally motivated approach to cognition to occur in the same sentential contexts, different emerged in the mid 1980s, apparently eschewing ‘clusters’ of words emerged, with each cluster repre- symbolic computation and modularity. senting a different category of word. Not surprisingly, the entire connectionist enterprise The Late 20th Century Emergence of came under intense critical scrutiny from the linguis- Connectionism: Statistical Approaches to tics and philosophy communities, not least because it Language appeared to reduce language to a system of statistical patterns, was fundamentally associationist, nonmod- In 1986, David Rumelhart and Jay McClelland pub- ular, and eschewed the explicit manipulation of sym- lished Parallel Distributed Processing. This edited bolic structures (because the internal representations volume described a range of connectionist,orneural that emerged as a result of the learning process were network, models of learning and cognition, and not symbolic in the traditional sense). Within the marked a ‘coming of age’ for connectionism. It was, context of the symbolic-connectionist debate there for many researchers in psycholinguistics, their first developed what became perhaps one of the longest introduction to a wide range of research in this emer- surviving disputes in contemporary psycholinguistics; ging field. Of particular interest were the facts that between those that believe that word formation (e.g., ‘knowledge’ in connectionist networks is encoded as the formation of ‘walked’ from ‘walk,’‘ran’ from patterns of connectivity distributed across the neural- ‘run,’ and ‘went’ from ‘go’) is driven by knowledge like units, and ‘processing’ is manifest as spreading of rules and exceptions to those rules, and those who patterns of activation. These networks can learn believe it is driven by statistical regularity (which can complex associative relations largely on the basis of apparently capture, in the right model, both the regu- simple associative learning principles (based primari- larly and irregularly formed words). The debate ly on work published in 1949 by Donald Hebb, a shows little sign of abating, even 20 years later. student of Lashley’s). Various algorithms exist to set Critics notwithstanding, statistical approaches to the ‘strengths’ of the connections between the units language (both with respect to its structure and its automatically, so that a given input pattern of activa- mental processing) are becoming more prevalent, tion across some set of units will spread through the with application to issues as diverse as the ‘discovery’ network and yield a desired output pattern across of words through the segmentation of the speech some other set of units. Indeed, multiple input- input, the emergence of grammatical categories, and output pairings can be learned by the same network. even the emergence of meaning as a consequence of Psycholinguistics: History 263 statistical dependencies between a word and its con- system a representation of the external world. The job text (cf. Wittgenstein’s views on the meaning of of the cognitive system is then to reconstruct, mental- words). Empirically also, the statistical approach ly, that external world. This reconstruction subse- has led to investigation of issues ranging from infants’ quently forms the basis for ‘commands’ sent to, for abilities to segment speech and to induce grammar- example, the motor system. Cognition thus medi- like rules to adult sentence processing. The reason ates between perception and action. An alternative that such approaches have proved so appealing is approach, termed ‘embodied cognition,’ is that cog- that statistics are agnostic as to the nature of the nition and action are encoded within the same repre- real-world objects over which the statistics are sentational medium. Cognition is thus rooted in the calculated – thus, the fundamentally same algorithm same motoric and sensory representations that sup- can be applied to sequences of phonemes, words, or port interaction with the external world. Or, put an- sentences. Their implementation within a neural net- other way, cognition is grounded in the same neural work is similarly agnostic – the same network and substrates that support sensory-motoric interaction the same algorithms that enable that network to with the external world. One consequence of this induce the appropriate statistics can be applied to view is that language, a component of cognition, many different domains. Connectionism opened up should, like the other components of cognition, be experience-based learning to a range of psychological studied in the context of (i) the interactions it causes domains, not just the linguistic domains. And experi- between the hearer and the world, and (ii) the neural ence-based learning was attractive not least because it substrates that support those interactions. Coinciden- required fewer assumptions about the existence of tally, the 1990s saw a boom in research into the neural innately specified domain-specific faculties (and in a substrate of language, in part due to the increased multi-authored volume published in 1996, Jeff Elman availability of neuroimaging technologies (predomi- teamed up with a variety of developmental psycholo- nantly PET and fMRI, with EEG and more recently gists to argue that connectionism was attractive pre- MEG also proving influential). It also saw increased cisely because it enabled a new perspective on how research into the relationship between language and innate constraints on learning and neural structure action. Taken together, these two streams of research might be an important component of human lan- provided increasing evidence for embodied cognition. guage acquisition (Elman, 1996)). With respect to imaging, a variety of studies Neural networks can be criticized for being (among demonstrated what Lichtheim had alluded to a cen- other things) too unconstrained – they can, in princi- tury earlier – that concepts are not represented in ple, do more than might be humanly possible – but some discrete location within the brain, but are the opposite criticism, that they are too small and do distributed across different regions. For example, not necessarily ‘scale up’ is another criticism that is words whose meanings implicate tool use (e.g., ‘ham- often heard. Neural networks as currently implemen- mer’) activate regions of the brain responsible for ted are just the ‘medium’ on which are offered up the controlling motoric action (during the use of the statistics. To misuse a common adage, the proof will tool) and other regions involved in the recognition be in the pudding, not in the plate that serves it up. of object form (during perception of the tool). Color There is little doubt, from the historical perspective, words (e.g., ‘yellow’) and words referring to non- that although the emergence of connectionism has manipulable artefacts (e.g., ‘house’) do not activate offered a powerful theoretical tool, its emergence motoric areas to the same extent, but they do activate has also polarized sections of the psycholinguistic regions close to those implicated in form perception community, between ‘connectionists’ on the one hand, and, for color words, color perception. Importantly, and ‘symbolists’ on the other. This polarization is not there is no single region that is primarily active; rath- unique to psycholinguistics, however, but pervades er, words and concepts activate complex patterns the study of cognition more broadly. And as if to of activity that are distributed and overlapping further muddy the theoretical waters, the beginning within different parts of the brain that are known of the 21st century has seen renewed interest in yet to have (other) motoric and sensory functions. The another (no less controversial) paradigm – one that meanings of (at least some) words do appear, then, to grounds language (and cognition) in action. be grounded in those neural substrates that support sensory-motoric interaction. The Early 21st Century and the Grounding About the same time that increased attention was of Language in Action and the Brain focusing on neuroimaging, new techniques for study- ing language and its effects on action were also being Traditional theories in cognition suppose that the job developed. One of these involved the monitoring of the perceptual system is to deliver to the cognitive of eye movements as participants listened to 264 Psycholinguistics: History commands to manipulate objects in front of them, or (e.g., language breakdown, normal language use, as they listened to descriptions of events that might ambiguity resolution, and so on), but also how we unfold within the scene before them (one can view study those behaviors (through studying aphasis, neu- language-mediated eye movements as central to the roimaging, language-mediated eye movements, and relationship between language and action, because so on). And we have still to see the full influences of eye movements signal overt shifts in attention, and connectionism, statistical learning, embodied cogni- because attention to something necessarily precedes tion, and the neuroscience of language. What we can (deliberate) action upon it). It was found that eye be sure of is that the boundaries between the study of movements were closely synchronized with processes language and the study of other aspects of cognition implicated in both spoken word recognition and sen- are wearing thinner (the eye movement research tence processing, and that much could be gleaned mentioned above is at the interface of language and about what kinds of information were recruited at vision, for example). No doubt there are already what point during a word or sentence in order to developments in ‘neighboring’ fields of study (e.g., interpret the unfolding language with respect to the the computational sciences and non-cognitive neuro- scene in front of the participant (it is not without sciences) that will also have an impact, but have yet to some irony that in L. N. Fowler’s famous Phrenology emerge as quantifiable influences on psycholinguis- bust, from about 1865, the faculty for language is tics. For example, researchers are already using located just below the left eye). Another technique computational techniques coupled with detailed neu- involved measuring motoric responses to different roanatomical research on the neural structure of the kinds of linguistic stimuli – for example, words or brain to attempt to understand the kinds of ‘computa- sentences referring to movements toward or away tion’ that distinct parts of the brain may be capable of. from the body were found to interfere with responses Such research promises greater understanding of the in a judgment task (e.g., ‘does this sentence make brain’s ability to learn, represent, and deploy lan- sense?’) that involved moving a finger toward or guage. And although the history of psycholinguistics away from a response button. A range of studies, is relevant to understanding where the field is today, some involving TMS (Transcranial Magnetic Stimu- perhaps of greater interest is where the field will be lation – a method for either temporarily stimulating tomorrow. or suppressing a part of the brain, such as parts of motor cortex) have confirmed this motoric compo- nent to language comprehension. Acknowledgments It is noteworthy, with respect to the embodiment This article is a substantially expanded version of a approach to cognition, that some of its basic tenets short, 1500-word section on the history of psycholin- have been around since the earliest days of (contem- guistics that appeared as part of a broader review of porary) psycholinguistics. Winograd’s SHRDLU pro- issues in psycholinguistics (Altmann, 2001). The gram, for example, viewed the meaning of a word reader is referred to this paper for in-depth review such as ‘place’ or ‘lift’ as that part of the program that of (some of) the topics that constitute the field of caused placing or lifting – language comprehension psycholinguistics. The author would like to thank within that program was grounded in sensory-motoric the publishers of this article for donating his fee representation – and as such, SHRDLU followed in to the charity Me´decins sans Frontie`res, and Silvia the Wittgenstinian tradition of treating meaning as Gennari for advice on topics ranging from Plato to use. Similarly, it is noteworthy that although most of neuroimaging. the neuroimaging of language has been carried out independently of theories of embodied cognition, much of the work converges on the same theme – See also: Cognitive Science: Overview; fMRI Studies of that aspects of language are represented in the same Language; Human Language Processing: Connectionist Models; Human Language Processing: Symbolic Models; representational substrates that control our sensory- Language, Visual Cognition and Motor Action; Modularity motoric interactions with the external world. of Mind and Language; Psycholinguistics: Overview; Sen- tence Processing. Epilogue And that, broadly speaking, is where the field is now. In the space available, it is impossible to docu- Bibliography ment all the trends that have influenced contempo- Altmann G T M (1997). The ascent of Babel: An explora- rary psycholinguistics, and which have influenced tion of language, mind, and understanding. Oxford: not just what kinds of language behavior we study Oxford University Press. Psycholinguistics: Overview 265

Altmann G T M (2001). ‘The language machine: Psycholin- Fodor J A (1983). Modularity of mind. Cambridge, Mass: guistics in review.’ British Journal of Psychology 92, MIT Press. 129–170. Marcus G F (2003). The algebraic mind. Cambridge, Mass: Bates E & Goodman J C (1999). ‘On the emergence of MIT Press/Bradford Books. grammar from the lexicon.’ In MacWhinney B (ed.) The Martin A & Chao L L. ‘Semantic memory and the brain: emergence of language. Mahwah, NJ: Lawrence Erlbaum structure and processes.’ Current Opinion in Neurobiol- Associates. 29–80. ogy 11, 194–201. Blumenthal A L (1970). Language and psychology: Histor- McClelland J L & Elman J L (1986). ‘The TRACE model of ical aspects of psycholinguistics. New York: John Wiley speech perception.’ Cognitive Psychology 18, 1–86. & Sons. Price C J (1999). ‘The functional anatomy of word compre- Bever T G (1970). ‘The cognitive basis for linguistic struc- hension and production.’ Trends in Cognitive Sciences tures.’ In Hayes J R (ed.) Cognition and the development 2(8), 281–288. of language. New York: John Wiley and Sons. Rumelhart D E & McClelland J L (eds.) (1986). Parallel Chomsky N (1957). Syntactic structures. Mouton. distributed processing: Explorations in the microstruc- Chomsky N (1959). ‘Review of Skinner’s Verbal Beha- ture of cognition. Cambridge, MA: MIT Press. viour.’ Language 35, 26–58. Saffran E M (2000). ‘Aphasia and the relationship of Elman J L (1990). ‘Finding structure in time.’ Cognitive language and brain.’ Seminars in Neurology 20(4), Science 14, 179–211. 409–418. Elman J L, Bates E A, Johnson M H, Karmiloff-Smith A, Skinner B F (1957). Verbal behavior. New York: Appleton- Parisi D & Plunkett K (1996). Rethinking innateness: Century-Crofts. A connectionist perspective on development. Cambridge, Winograd T (1972). ‘Understanding natural language.’ Mass: MIT Press/Bradford Books. Cognitive Psychology 3(1), 1–191.

Psycholinguistics: Overview A Anderson, University of Glasgow, Glasgow, UK people accomplish various information-processing ß 2006 Elsevier Ltd. All rights reserved. tasks. In the field of psycholinguistics this means a concern with the cognitive processes by which a string of sounds in an utterance, or marks on a page, Emergence of Psycholinguistics in the are processed to identify individual words and sen- Late 1950s and 1960s from the tences, and how this emerging structure becomes Chomskyan Revolution mentally represented as a meaningful concept. The goal of this process is to derive models that account Although the study of language has been part of for how people achieve this so rapidly and success- psychology from its earliest years, including for fully, given what we know about the general limita- example in the work of Wilhelm Wundt, the father of tions of human cognitive processing. To oversimplify: psychology, a distinct field of psycholinguistics the psychologist is concerned with how the linguistic emerged in the late 1950s largely in response to the units are processed and represented; the linguist is impact of Chomsky. In the preceding decades, notably concerned with the description of the structures that in the United States, psychology had been dominated emerge from any such processes. by the behaviorist approach of researchers such as B. F. Skinner. They treated language as a form of verbal Research Topics in the Early Years of behavior, which, like all other behavior, they believed Psycholinguistics was governed by simple stimulus–response associa- tions. Chomsky demonstrated the shortcomings of the In its early years psycholinguistics reflected the con- behaviorist approach in explaining the productivity of cerns of linguistics and the central role of syntax. language and its complexity, and his work, notably Psychologists such as Miller and Isard (1963) showed Syntactic structures (1957), provided a major impetus that the syntax influences the way people interpret for a new kind of psychological investigation of lan- sentences, and even how many words people can guage. This was driven by an interest in the mental remember from a string of words that make no representation of language in general and syntactic sense. More words are remembered from a ‘sentence’ structures in particular (see Psycholinguistics: History). like ‘Accidents carry honey between the house’ than Psychology since the demise of behaviorism has from strings with no syntactic structure such as ‘On again been concerned with understanding the way that trains hive elephants the simplify.’ The focus on 266 Psycholinguistics: Overview syntax and its importance in language processing led listeners is a major source of potential ambiguities of many psycholinguists to try to test the psychological interpretation. Many sentences in a language have reality of Chomsky’s theory of generative grammar. potentially more than one interpretation as they Experiments were designed to explore the notion that are processed, yet the reader or listener is usually when people process sentences, what they are doing is not aware of any problem in arriving at one clear retrieving the deep syntactic structure as described in interpretation of a sentence. Chomsky’s transformational grammar. So initially, a The cognitive architecture that underpins such an number of studies seemed to show that the relative achievement has been a source of much debate in ease or difficulty with which a reader or listener could psycholinguistics. Some researchers have held that in process a given sentence was directly related to its the frequent cases where more than one analysis of a syntactic complexity. Chomsky’s kernel sentences, sentence is possible, the reader or listener computes equivalent to active affirmative declarative sentences, all possible analyses in parallel. The difficulty of gar- were recalled most easily and processed most quickly, den path sentences has led others to propose serial while sentences including one or more transforma- models, where it is assumed that a single analysis is tions such as negative or passive forms were more computed and corrected later if this is needed. Models difficult to process. These kinds of study led to the have been proposed to account for the empirical evi- so-called derivational theory of complexity. This was dence on sentence-processing difficulties. These often superseded as it became clear that the results of many involve various versions of parallel analyses, where of the studies that apparently provided support for candidate analyses are only active for a brief period of this purely syntactic view of how people process sen- time or are ranked according to frequency in the tences could also be accounted for by the influence language or plausibility with the context. One of the of semantic factors. Although sentence processing most influential of these accounts is the constraint- remains one of the most significant topics in psycho- based model of MacDonald et al. (1994). The weight- linguistics, the range of language phenomena that are ings attached to each candidate analysis are based on studied has broadened considerably since the early the frequency of a syntactic structure in the language, 1960s. The assumption that the role of psycholinguis- the plausibility of the words in the sentence to their tics is to demonstrate the psychological validity of any assigned syntactic roles, etc. So this model is an ex- particular syntactic theory has also been overtaken. ample of an interactive model where many different sorts of information, syntactic, semantic, pragmatic, Models of Sentence Comprehension contextual, and frequency, can all play simultaneous roles by activating alternative interpretations of the Many of the models of sentence comprehension that incoming linguistic information. In models of this have been developed in psycholinguistics try to eluci- type, semantic factors can override syntactic proces- date the cognitive processes that are involved when sing biases. This model is attractive to psychologists a reader or listener interprets a sentence. There is for a number of reasons. It is amenable to modeling considerable experimental evidence that sentence by connectionist approaches (see Cognitive Science: comprehension is incremental, that is, that an inter- Overview) and it avoids the problem of having to pretation is built up on a moment-by-moment basis base cognitive processes on syntactic rules, which from the incoming linguistic information. The evi- for psychologists appear to change arbitrarily with dence for this incremental processing is particularly changes in linguistic theories. striking in the way listeners recognize spoken words In contrast, one of the other most influential mod- (see below), which are often identified before all the els of sentence processing is the garden path model acoustic information has been heard. (Frazier, 1979). This uses only syntactic principles in Even in written language processing we have clear its initial stage. An analysis is computed based on two signs of the incremental nature of linguistic proces- syntactic preferences, the most important being the sing. This is illustrated by the difficulties most readers principle of minimal attachment, the other being the have with sentences like the following: ‘The horse principle of late closure. The first principle means raced past the barn fell’ (Bever, 1970). This is known that the parsing of the sentence that produces the sim- as a garden path sentence, because nearly all readers plest parse tree, with fewest nodes, takes precedence. interpret this while they are reading, with ‘raced’ as So a sentence like ‘Mary watched the man with the an active verb and so expect the sentence to end after binoculars’ is usually interpreted to mean that Mary ‘barn.’ They do not realize that the sentence could (not the man) was using binoculars. According to have an equivalent interpretation to ‘The horse that Frazier’s interpretation of phrase structure rules, this raced past the barn, fell.’ The incremental way interpretation involves one node fewer than the alter- that sentences seem to be interpreted by readers or native and so demonstrates the principle of minimal Psycholinguistics: Overview 267 attachment in action. This principle takes precedence influenced how quickly the readers recovered from over the principle of late closure. This is the prefer- their wrong analysis of the syntax of the sentence. ence to attach incoming materials to the current Clifton et al. (2003) stressed that the key thing for phrase or clause. This latter principle is used to ex- psycholinguistic models of sentence processing is not plain the preference for interpreting sentences like whether the data on processing reduced relative ‘John said he will leave this morning’ to mean that clauses support garden path or constraint-based mod- the phrase ‘this morning’ relates to the verb ‘leave,’ els. They claim that the important goal is to develop not the verb ‘said.’ parsing models that deal both with the task of As part of the ongoing debate about the adequacy creating structure and evaluating the structure that of different models of the parsing process there has is created. been an active discussion in the experimental litera- Although there have been numerous studies of sen- ture over several years about the extent to which tence processing conducted by psycholinguists over semantic factors can override or guide the analysis the decades, the vast majority of these have focused of syntactic structure. This has been explored in sev- on how readers interpret written sentences. A few eral studies focusing on the ease or difficulty with studies have tackled the issue of how listeners use the which sentences containing reduced relative clauses cues in spoken language during parsing. Minimal can be processed. Several studies have tested how attachment strategy can be shown to be overcome people interpret sets of sentences like the following: by the prosodic cues in real spoken sentences. Simi- larly, the principle of late closure, which can cause (1a) The defendant examined by the lawyer turned syntactic ambiguities in written sentences, can be less out to be unreliable. (1b) The defendant that was examined by the lawyer problematical in spoken materials because of clear turned out to be unreliable. prosodic cues to the intended interpretation. Even the apparent errors in spoken sentences, such as the (2a) The evidence examined by the lawyer turned out disfluency ‘uh,’ can have an impact during the parsing to be unreliable. process. Bailey and Ferreira (2003) presented sets of (2b) The evidence that was examined by the lawyer turned out to be unreliable. sentences to listeners such as the following: Sentences like (1a), which contain a reduced relative (3a) Sandra bumped into the busboy and the uh uh waiter told her to be careful. clause, are more difficult to process than their equiv- (3b) Sandra bumped into the busboy and the waiter alent full relative clause (1b). Readers initially treat uh uh told her to be careful. ‘examined’ as a main verb whose subject is ‘the de- fendant.’ They then have to reanalyze this garden These sentences are ambiguous up to the point when path when they reach the phrase ‘by the lawyer.’ the listeners hear ‘told.’‘Sandra’ could have bumped The argument is the extent to which structurally sim- into ‘the busboy’ or ‘the busboy and the waiter.’ Yet ilar sentences (e.g., [2a]) cause readers to have equiva- the listeners who heard the materials most often inter- lent processing problems. This is what might be preted sentences like (3a) to mean that ‘the waiter’ expected from a purely syntactic view of parsing. In was the subject of a new clause, i.e., that it was the contrast, in an interactive constraint-based model, the subject of the verb ‘told.’ This shows that disfluencies semantic implausibility of interpreting an inanimate can systematically influence the way listeners parse noun such as ‘evidence’ as the subject of the verb incoming sentences. The same effect was observed ‘examined’ should protect the reader from the need when the interruption was not a disfluency but an to reanalyze an initial incorrect syntactic structure. environmental sound such as a telephone ringing. Trueswell et al. (1994) seemed to show just such a pattern. This was considered powerful evidence in Speech Production and Speech Errors support of interactive constraint-based models of sen- tence processing. More recently, Clifton et al. (2003) These studies represent a welcome aspect of the have challenged the evidence that semantic factors broadening of the psycholinguistic research agenda override syntactic processing in the initial stages of to include more consideration to the production and parsing. They used more sophisticated techniques for comprehension of spoken as well as written language. monitoring and analyzing eye movements to deter- The study of speech disfluencies is one part of this. mine the processing difficulties experienced by read- Speech disfluencies encompass a range of phenom- ers. They found that reduced relative clauses caused ena, including pauses in speech such as silences, filled disruption to processing, irrespective of the semantic pauses, and fillers such as ‘uh’ and ‘um,’ as well as plausibility of the relationship between the apparent speech errors such as slips of the tongue, spoonerisms, subject and main verb. Semantic factors, however, and malapropisms. When we speak we aim to 268 Psycholinguistics: Overview produce a grammatically well-formed utterance with models of production. Errors of word substitution no noticeable hesitations. Yet this ideal delivery can- usually involve words that are a phrase apart. Yet not always be achieved. It is estimated that around sound errors seem to relate neighboring words. This 5% of words in speech are disfluent in some way. Yet sort of evidence has been used by several researchers these disfluencies are not random in their patterns of to develop general models of the speech production occurrence. Even the similar-sounding disfluencies process (see Speech Errors: Psycholinguistic Ap- ‘uh’ and ‘um’ have been shown to have systematic proach for more details). contexts of use. Speakers use ‘uh’ before a short delay in their speech production but use ‘um’ before a more Speech Recognition significant delay. Speakers seem to become disfluent because they are experiencing some kind of problem Psycholinguists have also been concerned with ex- in planning and producing their utterance. Speakers ploring the processes involved in speech perception. have been found to pause more before unpredictable Jusczyk and Luce (2002) summarized over 50 years of words, suggesting they might be experiencing word- research on this topic. They described key research finding difficulties. Speakers also pause more at issues in the domain as understanding invariance, the start of an intonation unit, which suggests that constancy, and perceptual units. In speech there are pausing is related to the speech-planning process (see no invariant acoustic features that map directly to Pauses and Hesitations: Psycholinguistic Approach). corresponding phonetic segments. The acoustic prop- Speech errors have also been studied by psycholin- erties of sounds vary widely depending on the sur- guists, who have classified them according to the rounding linguistic context. To make matters even assumed units of processing and types of mechanism more complex for the listener, there is also wide involved in their production. For example, speech variability in the way phonetic segments are produced errors can relate to the phonemic features of the by different speakers depending on age, sex, and indi- word, the syllabic structure, or the phrase or sentence vidual speaker characteristics. It is not even easy to being produced. This is illustrated in one of the fa- determine what are the basic perceptual units that mous errors reportedly produced by Dr Spooner in listeners use to recognize speech. Some research stud- the 19th century when rebuking one of his students: ies seem to suggest the phoneme as the basic building ‘you have tasted the whole worm’ when he presum- block of perception while others show advantages of ably intended to say ‘you have wasted the whole the syllable over the phoneme. Conversational speech term.’ Here the initial phonemes are swapped but is therefore a very variable signal that does not even the rest of the morphemic and syntactic structure of provide clear cues to boundaries between words. Yet the target utterance is preserved. Errors of this type understanding words in speech is an effortless and became known as ‘spoonerisms’ as a result. successful process for listeners with normal hearing. The kinds of error that occur can tell us a good deal How is it done? about how speech is produced. Speech errors are very One of the key research challenges for psycholin- varied. They can reflect many different linguistic guists was therefore to produce models of spoken levels. Errors can involve the sounds of the words word recognition. One of the most influential models involved, for example saying ‘the lust list’ for ‘the is the Cohort Model (Marslen-Wilson and Welsh, lush list.’ They can relate to the intended words in a 1978; Marslen-Wilson, 1989). In this account, when phrase, ‘the pin of a head’ being said in place of ‘the a listener hears an initial sound, all the words known head of a pin.’ Errors may also focus on the semantic to start with that sound become activated. This co- relations of the intended words, so a speaker may hort of candidate words is gradually whittled down produce ‘I like berries with my fruit’ rather than to a single word, as more acoustic information is ‘I like berries with my cereal,’ among many other processed and candidate words are eliminated. An forms of errors. important feature of the model is the uniqueness Some types of errors, however, do not occur and point. This occurs when the listener has heard enough these patterns of occurrence and nonoccurrence have acoustic information to reduce the cohort to a single been used to help understand the speech production candidate, i.e., there are no other words known to the process. Content and function words are not substi- listener with that particular sequence of phonemes. tuted for one another and indeed substitute words are Syntactic and semantic context from the surrounding usually the same part of speech as the intended target discourse can also play a role in rejecting potential word. When the wrong sound is produced, however, candidate words. In later versions of the model, this the substituted phoneme seems to have no grammati- can only occur after the uniqueness point, though in cal relation with the intended target sound. The earlier versions, context could also be used to elimi- spacing between the errors is also informative for nate possible words before the uniqueness point. The Psycholinguistics: Overview 269 recognition point occurs when a single item remains computational model it has the virtues of specificity. in the cohort. This may be before the end of a word. Trace models can be built and simulations run and One of the strengths of the model is the way it can then compared with the experimental data from account for the speed with which spoken word recog- human listeners. These comparisons have generally nition often occurs, often occurring before the word shown that Trace can cope with some of the problems offset. In later versions of the model the activation of of variability in production of features and phonemes, candidate words is a graded process. Candidate can account for the context effects on spoken word words are not completely eliminated as acoustic in- recognition that have been reported in the literature, formation accumulates, but rather they have their and can cope with the kind of degraded acoustic input activation level reduced. This can rise again if later that is so typical of real conversational speech. The acoustic information matches a rejected candidate main criticisms that have been leveled at Trace con- word. This is important, as one of the problems cern the large role that context is given, which may be of conversational speech is that individual words or an overstatement of how this operates in human rec- phonemes are often not articulated clearly and ognition. The other limitation is the less than elegant the listener must be able to recover and recognize way that the time course of speech recognition is words whose initial sounds were mispronounced or modeled, with duplication of the levels and nodes misheard in a noisy environment. over successive time periods. The main competitors to speech recognition mod- Other connectionist models have also been devel- els are various connectionist accounts, notably Trace oped. Despite the competing architectures of the (McLelland and Elman, 1986). This is a connectionist models, there is general agreement on key aspects of or parallel distributed processing (PDP) account of the speech recognition process. This involves activa- spoken language processing. In such models research- tion of multiple candidate words, followed by com- ers were attempting to produce computational mod- petition among those known lexical items that share els of language processes which were inspired by a similar sound profile; these processes have to be what was known about the structure and processes able to cope with less than ideal acoustic input of the human brain (see Cognitive Science: Over- and deliver perceptions of words very rapidly (for view). In spoken word recognition models, this further details see Speech Recognition: Psychology meant that the feature units were simple units Approaches). with dense series of interconnections, which pro- cessed by sending many messages in parallel. Such Discourse Processing structures were developed as analogous to the neural architecture of the brain with its many nerves and The gradual broadening of the psycholinguistic re- interconnections. search agenda has not been limited to the inclusion The Trace model has three layers of units, of speech alongside written language. Psycholinguists corresponding at the lowest level to features, then at have also shown a growing interest in the processes the next layer to individual phonemes, then at the top involved in the interpretation not just of single sen- layer to complete words. All have dense arrays of tences, but of complete texts. One of the first psychol- interconnections between them, which like nerve ogists to explore the complexities of this process was pathways can be excitatory or inhibitory. The connec- Bartlett (1932). In his research on the way people tions between levels are excitatory and connections remembered and reproduced stories that they had within levels are inhibitory. Connections between heard, he highlighted key research themes in discourse levels operate in both directions, so both top-down processing that are still current research topics. Bart- and bottom-up processing can occur. The connectivity lett noted how quickly the surface form of the story is of Trace means that evidence is boosted and plausible lost and an individual’s own interpretation of the text hypotheses about the possible words emerge strongly. is what is remembered. He introduced the concept of So a feature such as voicing will energize the voice a ‘schema,’ which was used when readers recalled a feature units. These will then transmit activation to narrative. This consisted of an organized set of infor- all the voiced phonemes at the phoneme level, which mation based on prior experiences that is used will in turn activate all the words that begin with in interpretation and recall. The interpretation of a these phonemes. At each level the activated units narrative that is retained consists of a mix of input inhibit competitors at the same level, so reducing from the text and from schemata. In some of Bartlett’s possible competitors. A word is recognized when in studies, British students listened to North American the end a single active unit remains. folktales. When asked to recall the stories accurately, Trace is a very interactive model. It gives context a strange narrative details relating to the activities of bigger role in recognition than the cohort model. As a ghosts were unconsciously altered and supplemented 270 Psycholinguistics: Overview by details from the participants’ world knowledge, so Although there are significant differences between that the recalled version of the stories became more the models, there are several agreed features of how coherent by conventional Western standards. discourse processing operates in terms of the way More recently, a growing body of psycholinguistic readers update their mental representations of a research has been addressing the challenges of how text, the way some information is held in the fore- readers build up a coherent mental representation to ground of processing while others is background, and create a sense of the narrative world. One of the key the way that certain inferences are drawn automati- concerns of psycholinguistic studies of text or dis- cally to maintain a coherent account of the text. For course processing is the problem of inferences. Since details of the various models of discourse processing Bartlett’s seminal studies, it has been known that that have been proposed (see Discourse Processing). readers expand on what is in the text by drawing inferences based on their knowledge of the world. Reading as a Developmental and But what are the time course and limits of such Educational Process inferencing? Many studies have been concerned with addressing such questions. Experimental studies have Before an interpretation of a written text can be shown that some inferences seem to be made auto- made, the words on the page have to be read. matically as we read a text while some are made later Although apparently effortless for the skilled adult to resolve apparent problems or inconsistencies. This reader, the processes of identifying letters, recogniz- was demonstrated in a study by Sanford and Garrod ing words, and thus distinguishing the meaning con- (1981), who presented readers with pairs of sentences veyed in even a simple sentence, are complex. One such as the following: key to unraveling how this is accomplished has been to study readers’ eye movements. These have been (4a) Mary was dressing the baby. shown to be very systematic and to consist of three (4b) The clothes were made of wool. main types: short forward movements of around (5a) Mary was putting the clothes on the baby. 6–9 letters called saccades, which last on average (5b) The clothes were made of wool. 20–50 ms; backward movements called regressions; The participants read the second sentence just as and the pauses or fixations when the readers’ eyes rest quickly in both cases. This suggests that verbs such on a word for around 250 ms. as ‘dress’ cause readers to automatically draw the Rayner and colleagues have shown there are con- inference concerning ‘clothes.’ In contrast Haviland sistent patterns in these movements (see e.g., Rayner, and Clark (1974) found that some inferences took a 1998). When reading more difficult texts, fixations small but significant amount of time during reading. grow longer, saccades grow shorter and regressions In their experiment they used pairs of sentences like become more frequent. Even within a single text, the following: readers will spend more time fixating relatively uncommon words compared to familiar ones. Eye (6a) Harry took the picnic things from the trunk. movements are designed to keep the middle of our (6b) The beer was warm. visual field, where our vision is best, aimed at new (7a) Harry took the beer from the trunk. areas of interest. However, we are able to distinguish (7b) The beer was warm. quite a lot of information within a single fixation. They found that sentences like (6b) took longer to A skilled adult reader of English will usually be able read than (7b), because the readers had to make the to identify 15 letters to the right of the fixation point backward inference that the beer was part of the but only three or four letters to the left. picnic supplies. When children begin to learn to read they fixate Researchers wished to determine the limits on the words for longer than skilled readers. They also have kinds of inferences which are made immediately and a shorter perceptual span than adults, which means automatically and which are made later. Clearly, that in a single fixation they are able to identify fewer there must be limits on the amount of background letters. Gradually these patterns approach those of knowledge readers activate and one of the research adults as reading proficiency increases. The typical goals is to understand what these limits are and the English reader’s asymmetric perceptual span starts cognitive processes which support this. Models have to appear in most young readers within a year of been developed that attempt to specify the way starting to learn to read. This seems to reflect the inferences are made and the way coherent mental left-to-right nature of reading English, as readers of representations are derived from texts. (For a fuller Hebrew, which is read right to left, show the opposite account of inferencing see Coherence: Psycholinguis- pattern in their perceptual spans. Like much of psy- tic Approach). cholinguistics, the study of skilled reading has largely Psycholinguistics: Overview 271 been the study of skilled English reading but more similar words. Within a bin, files are organized by recently interesting studies have been conducted on word frequency. The details of the model are de- other languages, notably on nonalphabetic scripts scribed to account for the observed features of the such as Chinese and Japanese. (see Reading Processes recognition process. For example, there are cross- in Adults for more details). references between master files that would support semantic priming. Despite these features it is not clear Visual Word Recognition that this kind of serial search model can convincingly account for the speed with which words are recog- One aspect of the reading process has received a great nized. It has also been criticized as being based on deal of research attention in psycholinguistics: the a rather dated analogy with the cognitive system as a process of visual word recognition. A whole set of digital computer rather than as a neural system. phenomena have been identified in the processes of Very recently, however, Murray and Forster (2004) recognizing written words. These include the process have published an account of a series of word recog- of priming. Words are recognized more quickly if they nition studies that are claimed to show strong support have been read previously. This is known as repetition for the serial search model. They claim that the struc- priming. Words are also recognized more quickly if a ture within bins, notably the rank ordering of words word of similar meaning has just been presented, so within a bin in terms of frequency, accounts for the ‘butter’ is recognized more quickly if ‘bread’ has just pattern of experimental results on word frequency been read. This is known as semantic priming. Priming effects more parsimoniously than alternative direct can also occur between words which do not seem to access models. have a direct semantic relationship, such as ‘music’ More popular models of word recognition involve and ‘kidney,’ which are linked by an intermediate direct access from the sensory information to the word ‘organ,’ which was not presented to the readers. lexical units. One of the most influential of the Other phenomena which researchers have identified early accounts of this sort was the Logogen Model in word recognition include the fact that words which (Morton, 1969). In this model, perceptual informa- are common in the language, such as ‘road,’ are recog- tion, either from visual or auditory analysis, feeds nized more quickly than similar words that are less directly into the logogen system. This consists of a common, such as ‘rend.’ This word frequency effect is series of units, logogens, which represent known a strong influence on recognition speeds, with even words. Logogens act as feature counters and when a fairly small differences in word frequency influencing logogen has accumulated sufficient evidence to reach reaction times. It is the most robust effect in studies of a threshold it fires. A word then becomes available to word recognition, appearing in many different studies the output buffer, is recognized, and can be articulat- using a wide variety of research methods. ed. The logogens receive input from the cognitive These and other phenomena about how words are system as well as the sensory input routes and so the recognized have been used to develop a variety of resting threshold of the logogen can be varied by, for models of word recognition. These can be grouped example, prior experience of a word, or from the into families of related accounts. Some of the pro- sentence or discourse context. Common words with posed models are direct access models, where percep- which an individual has had a lot of prior experience tual information goes straight to feature counters or will have a lower threshold and will fire with less units. Other accounts propose that perceptual infor- sensory input and hence be recognized more quickly. mation is used to trigger a search through the mental Similarly, words that are highly predictable from lexicon, that is, the stored representation of all the context will also have their thresholds raised and so words known to the reader. will be recognized rapidly. One of the best-known serial search models The Logogen Model has been used as the basis of was proposed by Forster (1976). In this account, the computational models of word recognition, notably perceptual input is used to build a representation of the interaction activation model (IAM) proposed by the word to be recognized which is then checked in McLelland and Rumelhart (1981). This was one of two stages by comparisons with a series of access the early connectionist or PDP accounts of language files, which are analogous to the cards in a library processing. Researchers were trying to produce com- index system. Once an input string is matched to an putational models of language processes which were access file it is then linked to the master files, analo- inspired by what was known about the structure and gous to the books on the shelf, which contain the full processes of the human brain (see Cognitive Science: lexical entries for each word. The files are organized Overview). In word recognition models, this meant to speed up the process of word recognition, with that the feature units were simple units with dense groups of files being arranged in bins that contain series of interconnections, which processed by sending 272 Psycholinguistics: Overview many messages in parallel. The IAM consisted of descriptions of even complex and abstract stimuli three layers of units, corresponding at the lowest become much shorter and more concise. The interac- level to the visual features of letters, then at the next tion between the speakers was found to be crucial in layer to individual letters, then at the top layer to this process. If this interaction was disrupted the complete words, all with dense arrays of interconnec- speakers were not able to reduce their descriptions tions between them, which like nerve pathways could to the same extent. be excitatory or inhibitory. The model was able to Later studies on referential dialogues were con- account for a wide range of experimental observa- ducted by Clark and colleagues. In one influential tions on word recognition. (For more details see paper, Clark and Wilkes-Gibbs (1986) developed a Word Recognition, Written.) collaborative model of dialogue. From studying many Models of this type have become very widely used pairs of speakers engaged in dialogues, they high- in psycholinguistics to explore a wide variety of lan- lighted the way speakers work together through guage processes. This trend towards testing computa- their contributions to dialogue to arrive at a shared tional models against the existing experimental and mutual way of referring to things they wish to literature in a way that can link to new insights discuss. This means that the speaker and the listener about the neural processes involved in language and both take responsibility for assuring that what cognitive processes is a popular approach in current has been said is mutually understood or ‘grounded’ psycholinguistics. before the dialogue proceeds. In this view of dialogue, speakers follow the princi- Dialogue and Gesture ple of collaborative effort and try to minimize their overall effort in arriving at an agreed description. In parallel to this concern with understanding the This is done iteratively over a number of turns of cognitive processes and the possible neurological speaking, often with each speaker contributing part architecture involved in language processing has of the description. This process makes use of the been a growing interest in language processing in opportunities and limitations of spoken dialogue in context. This means not only the growth in studies a way that highlights its differences from written of how extended texts are processed but also a devel- language. In text, the writer can take as much time oping field of psycholinguistic studies of interactive as is needed to produce a description that she thinks language use. These studies have focused on language the reader will understand. In dialogue, however, the in its natural setting, interactive dialogue. speaker is under time pressure, as conversation rarely Dialogue represents the most ubiquitous form of allows long gaps in which to plan an utterance. The language use. As young children, we learn to use speaker therefore has to produce an immediate de- language through dialogues with our parents and scription and may not be able to retrieve the most caregivers. Even educated adults spend a great deal appropriate description or judge what way of describ- of their time in conversations with family, friends, ing a referent will be interpretable by the listener. The colleagues. Spoken dialogue is still used to obtain advantage of dialogue, however, is that the speaker many forms of goods and services. In the many non- and the listener can work together to refine or clarify literate cultures in the world dialogue is the main or the speaker’s initial description till they both share a only form of linguistic interaction. common understanding. Yet till recently dialogue received rather little The highly collaborative nature of this process is research attention in psycholinguistics. One of the reinforced in two key studies by Clark and colleagues. challenges for psycholinguists who wished to study Clark and Schaefer (1987) showed how contributing dialogue was to derive methods of exploring the to dialogue is characterized by two phases. Each time phenomenon which would produce testable and a speaker wishes to make a contribution, they pro- generalizable research questions and findings. One duce a stretch of speech that is the content of what experimental method which has been used to allow they wish to contribute. This is called the presenta- the study of comparable dialogues from many pairs of tion phase. They then require an acceptance phase, in speakers is the referential communication paradigm, which their listener gives evidence of understanding developed by Krauss and Weinheimer (1964). In this, the previous contribution. So the dialogue involves pairs of speakers are presented with an array of cards two activities: content specification and grounding, depicting abstract shapes. The speakers have to inter- that is attempting to ensure that both speakers un- act to determine which card a speaker is referring to derstand the content sufficiently for their current at a given point in the dialogue. These early studies conversational purposes. revealed key aspects of dialogue, including the way The importance of the ability to actively contribute that over the course of a dialogue, the lengths of to the dialogue interaction was demonstrated in a Psycholinguistics: Overview 273 study Schober and Clark (1989). Using the referential The important role of visual signals in the percep- paradigm they had pairs of speakers complete the tion of speech and how these are integrated with task. An additional participant was included in each acoustic information is a fascinating research area. interaction who overheard everything that was said This was highlighted in a seminal study by McGurk but did not take part in the dialogue. This overhearer and MacDonald (1976). They demonstrated that had a much harder time trying to identify the when a listener hears a phoneme such as ‘ba’ while intended referents of the descriptions than the conver- watching a face mouthing ‘ga,’ the sound which is sational participant, despite having the same pictures heard is a fusion ‘da.’ This is a powerful illusion that and having heard everything that was said. The sug- occurs even with knowledgeable listeners/viewers gested explanation is that the descriptions were not and has been demonstrated with young babies (see grounded for the overhearers. They had no chance to Audio-visual Speech Processing). collaborate and ensure that they understood each The role of visual signals in the production and description as it emerged during the dialogue. comprehension of more extended stretches of dis- So there is clear evidence that dialogue is an inter- course has also been the subject of considerable active and collaborative process that involves speak- study. From studies of conversation and storytelling, ers and listeners attempting to cooperate and achieve the important role of gestures and their relationship mutual understanding. The detailed mechanisms that to the accompanying speech has been established. For underpin these general processes of adaptation to the some kinds of gestures there is a close temporal rela- interlocutor are now the focus of a good deal of tionship with the accompanying speech. Listeners psycholinguistic study. There is some controversy also seem to fuse information presented visually and over the extent to which speakers are able to adjust verbally. If they are told a story by a speaker who and adapt their output to their listener’s needs. In uses speech and gesture and are then asked to retell terms of the forms of referring expressions chosen the story later, information originally presented by speakers there is evidence of adjustment to the by gesture, such as the speed of an action or the listener’s general level of knowledge of the domain. manner of leaving, is often relayed in speech and When it comes to adjusting the intelligibility of their vice versa (for more information Gesture and articulation, speakers seem to be largely egocentric. Communication). They reduce the clarity of their word production in terms of what is familiar to them as speakers rather Future Directions in Psycholinguistics than modeling their listeners’ needs. The time course of adaptation is also debated, with some studies Several trends seem apparent in psycholinguistics. showing speakers initially produce utterances from Some of these seem to be the result of improvements their own perspective but later monitor their listeners and developments in the research methods available and adapt. Other studies show listener adaptations to psycholinguistics. One is an increased interest in from the start of speaking (for more details see Dia- detailed investigations of language in richer, more logue and Interaction). naturalistic, contexts. New research techniques such As language processing is studied in more natural as improved methods of tracking a speaker or listen- contexts of use, it becomes clear that speakers and er’s eye movements mean, for example, that studies listeners do not just communicate using the verbal of dialogue, or the relationship of a speaker’s produc- channel. Visual signals from the mouth, face, hands, tion to the surrounding context, can be studied with and eyes are all important features of communication. the precision that used to be only possible in studies of Researchers have begun to explore the way the isolated word recognition or sentence processing. visual channel is used by speakers and listeners Improvements in the ease and accessibility of and the relationship between verbal and visual various brain-imaging techniques mean that these signals. are being used not only as contribution to our More generally, in a dialogue, what we say, how understanding of the neural substrate of different much we say, and even the clarity of the way the language processes. Techniques such as ERP (event- words will be spoken have all been shown to change related brain potentials) can now be used more and when speakers do or do not have access to visual more as means to explore the precise time course of signals. So speakers who can see one another need language processing. Newer techniques such as MEG to say less to complete a task, use more gestures, can (magneto-encephalography) are beginning to offer exchange turns of speaking more smoothly, and psycholinguists not just good information about the articulate their words less than when they cannot temporal patterns of language processing but also see one another. detailed information about the location of associated 274 Psycholinguistics: Overview brain activity. (For more information see Psycholin- Chomsky N (1957). Syntactic structures. The Hague: guistic Research Methods.) The growing interest in Mouton. the neural substrates which support language Clark H H & Schaefer E (1989). ‘Contributing to dis- processing has received a major boost from the course.’ Cognitive Science 13, 259–294. development of these new forms of brain imaging. Clark H H & Wilkes-Gibbs D (1986). ‘Referring as a collaborative process.’ Cognition 22, 1–39. The interest in how to build neurologically plausi- Clifton C, Traxler M & Mohamed M (2003). ‘The use of ble models was of course one of the drivers behind the thematic role information in parsing: syntactic processing expansion over the last 20 years in connectionist autonomy revisited.’ Journal of Memory and Language models of language. These have made a major contri- 49, 317–334. bution to our understanding of how a wide variety of Ferreira F & Clifton C (1986). ‘The independence of syn- language processes, such as spoken or written word tactic processing.’ Journal of Memory and Language 25, recognition, might operate and be learned. The chal- 348–368. lenge in the future will be to see whether connectionist Forster K I (1976). ‘Accessing the mental lexicon.’ In Wales models can be implemented for more extensive lan- R W E (ed.) New approaches to language mechanisms. guage processing, such as text comprehension or con- Amsterdam: North Holland. 257–287. tribution to dialogues. The way such models can or Frazier L (1979). On comprehending sentences: syntactic parsing strategies. West Bend: Indiana Universities cannot be scaled up to simulate more complex lan- Linguistic Club. guage processing will be one of the key challenges for Haviland S E & Clark H H (1974). ‘What’s new? Acquir- the next few years. ing new information as a process in comprehension.’ In the future psycholinguistics will also need to Journal of Verbal Learning and Verbal Behavior 13, address its undoubted Anglocentric bias. The vast 512–521. majority of studies of language processing are in fact Jusczyk P & Luce P (2002). ‘Speech perception and spoken studies of English language processing. In a number word recognition: past and present.’ Ear and Hearing 23, of areas a few studies are emerging which consider 2–40. other languages but this effort needs to be greatly Krauss R & Weinheimer S (1964). ‘Changes in reference increased. Over the last 50 years psycholinguistics phrases as a function of frequency of use in social inter- has expanded dramatically and made considerable actions: a preliminary study.’ Psychonomic Science 1, 113–114. progress in understanding a wide variety of language MacDonald M, Pearlmutter N & Seidenberg M (1994). processes. With new research techniques and a more ‘Lexical nature of syntactic ambiguity resolution.’ balanced research portfolio in terms of the languages Psychological Review 101, 676–703. studied and the research efforts applied to production Marslen-Wilson W (1989). ‘Access and integration: pro- as well as comprehension, spoken as well as written jecting sound onto meaning.’ In Marslen-Wilson W language, future progress seems assured. (ed.) Lexical access and representation. Cambridge, MA: Bradford. 3–24. See also: Audio-visual Speech Processing; Cognitive Sci- Marslen-Wilson W & Welch A (1978). ‘Processing inter- ence: Overview; Coherence: Psycholinguistic Approach; actions and lexical access during word recognition in Dialogue and Interaction; Discourse Processing; Gesture continuous speech.’ Cognitive Psychology 10, 29–63. and Communication; Pauses and Hesitations: Psycholin- McClelland J & Elman J (1986). ‘The TRACE model of guistic Approach; Psycholinguistic Research Methods; speech perception.’ Cognitive Psychology 18, 1–86. Psycholinguistics: History; Reading Processes in Adults; McClelland J & Rumelhart D (1981). ‘An interactive acti- Sentence Processing; Speech Errors: Psycholinguistic vation model of context effects in letter perception 1: An Approach; Speech Production; Speech Recognition: Psy- account of the basic findings.’ Psychological Review 88, chology Approaches; Word Recognition, Written. 375–407. McGurk H & MacDonald J (1976). ‘Hearing lips and seeing voices.’ Nature 264, 746–748. Bibliography Miller G & Isard S (1963). ‘Some perceptual consequences of linguistic rules.’ Journal of Verbal Learning and Verbal Bailey K & Ferreira F (2003). ‘Disfluencies affect the pars- Behavior 2, 217–228. ing of garden-path sentences.’ Journal of Memory and Morton J (1969). ‘Interaction of information in word Language 49, 183–200. recognition.’ Psychological Review 76, 165–178. Bartlett F C (1932). Remembering: a study in experimental Murray W & Forster K (2004). ‘Serial mechanisms in lexi- and social psychology. Cambridge: Cambridge University cal access: the rank hypothesis.’ Psychological Review Press. 111, 721–756. Bever T G (1970). ‘The cognitive basis for linguistic struc- Rayner K (1998). ‘Eye movements in reading and infor- ture.’ In Hayes J R (ed.) Cognition and the development mation processing: 20 years of research.’ Psychological of language. New York: John Wiley. 270–362. Bulletin 124, 372–422. Psychosis and Language 275

Sanford A J & Garrod S (1981). Understanding written Trueswell J, Tannenhaus M & Garnsey S (1994). ‘Semantic language. Chichester, England: John Wiley. influences on parsing: use of thematic role information Schober M & Clark H H (1989). ‘Understanding by addres- in syntactic disambiguation.’ Journal of Memory and sees and overhearers.’ Cognitive Psychology 21, 211–232. Language 33, 285–318.

Psychosis and Language M Simard, Universite´ Laval, Que´ bec City, Canada observation of the disorder, the concept of thought Y Turgeon, Campbellton Regional Hospital, disorder was transformed into problems of language Campbellton, Canada and communication behavior in the work of Andreasen ß 2006 Elsevier Ltd. All rights reserved. (1979) and into disorganized speech in the criteria for schizophrenia in the Diagnostic and statistical manual of mental disorders (4th edition) (DSM-IV) (APA, Several psychiatric disorders are included under the 1994). broad definition of psychosis. The most common dis- In the mid-20th century, Kurt Schneider (1957) orders are schizophrenia (which is prototypical of the described the primary disorders of the experience of psychoses), schizophreniform disorder, schizoaffec- thought representing the current core features of the tive disorder, and delusional disorder (APA, 1994). positive symptoms. This characterization is known as These other psychotic disorders share some distinc- the Schneiderian First-Rank symptoms, which in- tive features with schizophrenia (e.g., a schizoaffec- clude hearing voices speaking one’s thoughts aloud, tive disorder is a disturbance in which a mood episode hearing voices arguing about oneself, hearing voices and characteristic symptoms of schizophrenia occur commenting on one’s actions as they are occurring, together). Furthermore, an array of psychiatric and having bodily sensations imposed from outside, attri- medical conditions also presents with psychotic attri- buting one’s feelings to external sources, experiencing butes and is often characterized by presenting typical one’s drives as originating from powerful outside symptoms of schizophrenia. forces, moving and acting as a result of external con- The contemporary consensual definitions of trols, having one’s thoughts withdrawn from the schizophrenia and most common psychotic disorders mind, having one’s thoughts inserted into one’s mind, of the Diagnostic and statistical manual have resulted broadcasting of thoughts, and attributing special per- from more than 100 years of conceptual work. The sonal significance to one’s perceptions. Schneider’s current criteria for schizophrenia derive from the symptoms were thought to be pathognomonic signs, early works of Emil Kraepelin (1899), Eugen Bleuler any of which was strongly indicative of a schizophrenic (1911), and Kurt Schneider (1957). disorder. Psychosis is not a unitary concept. The ‘psychotic’ Background History of the Contemporary designation is often applied to symptom presentation Diagnostic Criteria for Schizophrenia, that may vary considerably, both biologically and Autism, and Asperger’s Syndrome behaviorally. The wide variety of psychiatric and medical conditions that present with ‘psychotic’ fea- Emil Kraepelin meticulously described the symptoms tures attests to the heterogeneity of psychotic disor- of schizophrenia, which he named dementia praecox, ders. It is now recognized that schizophrenia is at referring to the disruption of emotional and cognitive least clinically a heterogeneous disorder (Ragland, features as well as to the deteriorating course of the 2003), and this characteristic necessarily has an im- disorder. He also tentatively proposed a pathophysio- pact on the presentation of language and speech logical localization, suggesting that abnormalities in impairments that may arise from that condition. the frontal lobe would be the substrate of problems The heterogeneity of autism is also so well ac- with reasoning and volition, whereas abnormalities in knowledged (Eigsti and Shapiro, 2003) that the term the temporal lobe would be the substrate of delusions autistic spectrum disorder (Rapin and Dunn, 2003) is and hallucinations (Kraepelin, 1899). frequently used to describe the different extents of In his conception of schizophrenia, Eugen Bleuler severe functional deficits in sociability, communica- emphasized the importance of a formal thought disor- tive language, imaginative play, and range of interests der (disorder of associations or ‘splitting’). To over- that principally characterize this condition as per the come the limitations of Bleuler’s approach in the DSM-IV (APA, 1994).