Cohort model

The cohort model is a model of the auditory word recognition, which was developed in the 80s of the 20th century by scientists at the American psychologist William Marslen - Wilson. It describes how to access individual words stored in the mental lexicon, while listening. Here, the one literal is filtered for sound from a set of potential candidates, which corresponds to the incoming acoustic information at best.

• 2.1 Shadowing experiments
• 2.2 Word - monitoring experiments
• 2.3 Conclusions from the results

• 3.1 Serial word recognition
• 3.2 Declaration of context effects
• 3.3 criticism

• 4.1 Frequency effect
• 4.2 Defective input
• 4.3 criticism

• 6.1 Primary Sources
• 6.2 secondary literature

Overview

Word recognition vs.. speech recognition

The cohort model is a model of the auditory word recognition, that is, trying to explain how access works on the data stored in the header information of individual words when listening to the language. It is about the question of how, it is concluded that the appropriate entry in the mental lexicon of the acoustic signals, which together make up a word. For these models initially not of interest, such as the meanings of individual words within the overall context of a spoken utterance arise. Rather, it 's all about the meaning of the words themselves, which is only specified by the context.

Here is an example: A ambiges word such as bank encompasses many different meanings (for example, river bank, credit institution, seating, ...). Which is it just meant in a spoken utterance, resulting in most cases from the linguistic and non- linguistic context. In theory, all meanings of a word in the same record are stored in the mental lexicon, as well as all meanings of the word bank. The actual meaning is opened up when promoted by the context ( in technical jargon: primed ) is. This also applies to morpho - syntactic processes. For example, initially also accessed banks on the entry for the lexeme Bank in the understanding of the word, then for the plural ending -en. In another, the word recognition but even irrelevant process, the two recognized lexemes are merged and the word in its specific meaning ( here: several banks, but not more seating, since the plural banks would be ) understood. The same is true for complex expressions like endocentric compound words (eg, a bank employee ) or -derived words (eg, workers from the entries for the verb stem work and the substantivierenden suffix - er). Also for speech recognition, but not for word recognition, which belongs to syntactic analysis, such as the question of whether the word bank, embedded in a sentence as the subject ( the bank is located in the city center ) or as an object ( The ATM is located in the bank) acts.

Is it generally comes to finding a concrete meaning of a word or a sentence, depending on the semantic, syntactic and linguistic context for models of speech recognition to models of word recognition limited only to finding the entry in the mental lexicon according to the to-understand words. The existence of such mental lexicon is assumed as a given. In summary, models of word recognition generally deal with the question of how to access this dictionary. With such models, it is initially irrelevant, the direct significance of the entry is set ultimately by the context. Thus, the word recognition is a part of the speech recognition, but not the same as her.

History and definition

In psycholinguistics, a distinction is roughly two kinds of models on auditory word recognition. First, there are the phonological approaches, on the other hand, the psycholinguistic. The phonological research approaches are historically older than the psycholinguistic, but both continue to exist side by side and are adopted by different scientists in any given issue.

Phonological approaches

The phonological approaches describe word recognition as a dynamic process that is being introduced in which, after receipt of the acoustic signal in the course of repeated recognition to this signal. This is known as bottom- up approaches, that is, word recognition is based only on the acoustic signal. In addition, these models are based on earlier abstraction. Thus, the incoming signal is quite early, broken down into discrete units, such as distinctive features. Also phonological properties of language play an important role in the recognition.

As examples, the motor theory, the Acoustic Invarianztheorie or the quantum theory of speech perception should be mentioned.

Psycholinguistic Approaches

The psycho -linguistic approaches, however, focus on the word segmentation and word recognition. Plays world knowledge in phonological approaches no or only a minor role, is used for the repair of phonetically mangled signals in the psycholinguistic approaches this. Thus, these models are top-down oriented, which means that existing knowledge, for example the structural properties of the known words, resorted in the recognition of words. The cohort model is one of the first psycholinguistic theories of auditory word recognition. Others include, for example, the TRACE model of McClelland and Elman, as well as its further developments Shortlist model or merge model, which rely in part on mechanisms of the cohort model.

Scientists led by William Marslen -Wilson developed the cohort model the early 80s of the 20th century based on a series of experiments, the results argued that the previous models could not explain or only with additional assumptions.

General operation

The model

Marslen -Wilson divided the auditory word recognition in three macro levels: access (access ), selection (English selection) and integration.

Under Access the conversion of acoustic signals is understood in features or sounds in the model. In the selection phase is the mechanism of the cohort formation (see below) the appropriate entry is selected in the mental lexicon and embedded in the integration phase in the appropriate semantic and syntactic context. The model alone does not make any statements about the importance of detecting the complex units of meaning as sentences or phrases.

Lexical access in the cohort model

The basic idea of the model is that the incoming acoustic signal (called the input) is serially broken when listening to spoken language in Phone. The listener recognizes the first sound of the word to be understood and opens a set of all in his mental lexicon stored entries that begin with that same sound. This amount of lexical entries is called a cohort. In the next step, the second volume of the word is analyzed. Of the first cohort, all the entries are now selected, the second volume corresponds to the recognized of the input. The other lexemes are removed from the cohort. In other words, only the lexemes are available that match still with the previously identified information of the input. This procedure is now repeated with the following sounds, until the word is clearly recognized. This is the case when the cohort contains only one entry. The diagram illustrates this general operation of the model using the example of the recognition the English word " trespass " ( to German: trespassing ).

The original cohort model was to explain the situation, context effects and the serial nature of auditory word recognition. After the weaknesses of the model frequency effects and dealing with a defective input were known, advanced Marslen -Wilson model in the mid-1980s. In the literature the terms Cohort I and Cohort II have been naturalized for these two stages of the model.

Experimental foundations

William Marslen -Wilson conducted a series of experiments that exhibited two important features of auditory word recognition. Firstly, the serial character and on the other hand the influence of contextual information on the word recognition. At the same time he was able to demonstrate weaknesses in the models previously existing on speech perception. From the results of these experiments, the cohort model is then grown up, which should explain the observed effects.

Shadowing experiments

The first experiments that carried out Marslen -Wilson, were so-called shadowing experiments. In these experiments, the experimenter reads a text, which must be repeated after as quickly as possible from the subject. With an average word length of 500 ms, there was a delay between the words of the experimenter and the repetition of the subject of 250 ms. This means that the subject was able to recognize and repeat a word, even before the experimenter had the word correctly pronounced. Minus the time it takes the articulation of the perceived words to complete, it is now believed that the process of recognizing a pure word vonstattengeht within about 200 ms. This corresponds to at normal speaking speed of a length of about two to three sounds ( phonemes ).

In this context, one speaks of the uniqueness and recognition points as above. The uniqueness point (also Diskriminationspunkt ) is the point at which a word is recognized without any doubt, so if there is no other word which is encoded by the same phoneme sequence to be recognized as that. This is then placed at the latest, when a new word begins. The recognition point, however, is the point from which the listener can say with great certainty what word he perceives, ie after about 200 ms.

In further such experiments the subject a text was read, but contained errors, in contrast to the first experiment. The position of the error varied within a word, the error at the beginning, could be placed in the middle or at the end of the word. The adjacent table shows some examples. This partially erroneous words were packaged in three different texts, the one was a normal text. The second text was semantically anomalous sentences, the individuals were grammatically correct, but gave no coherent sense. The third text was semantically and syntactically anomalous, ie an incoherent string of words.

In these experiments, the focus was on the restoration of the erroneous words. Under restoration is understood, when the subject an erroneous word can still accurately reproduced and does not recognize it as being defective. It turns out that the subjects restore faulty words most likely when the error at the end of the word and under normal conditions occurred. Did the error, however, at the beginning of a word or in abnormal contexts, they were almost always recognized as such and not restored the corresponding words.

On the one hand, these observations speak for a serial character of the auditory word recognition, which explains the restoration at later time points. On the other side, such experiments provide information on the role of context in the detection of erroneous words, which can be through the restoration in normal contexts and noting the error in abnormal contexts show.

Word - monitoring experiments

A third series of experiments was so-called word - monitoring experiments. The test person gets it at short intervals a series of words presented acoustically via headphones and has the task of pressing a button located in front of him, when she hears a predetermined word. Measured the reaction time, ie the time between the beginning of the word to be recognized and pressing the button. The particular word was able to appear in three different contexts. Either by a semantically relatierten, after a syntactically relatierten or after a unrelatierten word. If the word to be recognized, for example, "Adler", then a semantically relatiertes word " bird" ( because bird is hyperonym to Adler ), a syntactically relatiertes word " the " ( " the eagle ") and a unrelatiertes word "blue" ( " blue "and" eagles " are neither semantically nor be in the form syntactically related ).

In these experiments, it was found that in a semantically relatierten Word that precedes the reaction time was the shortest. On syntactically relatierten predecessor word, the reaction time was significantly longer, longest at unrelatierten predecessor words.

Conclusions from the results

In addition to the detection of the serial nature of the auditory word recognition, the experiments in addition demonstrated the so-called context - effect.

In certain linguistic contexts, the occurrence of certain classes of words is more likely than the other. Thus, it is relatively unlikely, for example, in languages ​​such as German or English, according to an article that follows a verb. In this case one speaks of the syntactic context of the specific word classes primes, so can be probable. Priming reveals himself in a shorter detection time as opposed to unprimed lexemes.

In addition to the syntax and the semantic context plays an important role in the speed of the text input. For example, it is unlikely that in a speech act of nuclear physics has on the subject, a technical term is used in the landscaping. Similar effects are also seen at the phonological level. Rhymes, for example, two consecutive words, the second is recognized faster. Not rhyme it on the other hand, the detection of the second word takes longer.

Cohort I

Serial word recognition

The serial nature of the auditory word recognition is explained in the cohort model as follows: When a word is spoken, it reaches the listener in the form of acoustic waves, with each sound of the word as a specific pattern of overlapping waves coded. A lexical access to the upstream cognitive module deduced from these waves now, what specific sound they each represent. Under serial is to be understood in this context, that at any time exactly one is transmitted sound. Accordingly, it is therefore not the case that at a time when two different sounds are transmitted from the speaker to the listener. The to-understand word is therefore considered sound for sound.

After detecting the first loud the first time accessing the mental lexicon, which is the modulus of the cognitive system, in which all the words are stored, which have been learned in the course of a lifetime. In this case, all entries are activated that begin with the detected sound. Under activation, is understood in psycholinguistics that is accessed an entry, including further information is retrieved. These are called also adopted in entry contains. This information can, for example,

• Be semantic in nature, ie which contain information about the meaning of the word;
• Morpho - syntactic his nature, ie information about whether it is for example, a verb or a noun if the word requires arguments, which genus it belongs to, etc.;
• Its phonological nature, which state, among other things, with what sounds the word is formed.

With the recognition of the following sounds all the entries are removed from the cohort who no longer fit the heard, that is, the activation is canceled, the retrieved information is "forgotten". At a certain point, the cohort is only an entry large, this is the point from which the word is recognized unequivocally. In the literature, this point is called the uniqueness point. In this first version of the cohort model Marslen - Wilsen follows an all-or - nothing approach, a lemma can be either activated or not activated, there are no gradations or differences in activation.

Marslen -Wilson made ​​this clear to the English word trɛs.pʌs ( trespass, trespassing dt ). The first thing the listener recognizes the phoneme [t ] followed by [r ], then [ ɛ ], etc. Already at the first sound a cohort is opened and mental. In this case, all words are capitalized, which are listed in the mental lexicon and also start with the sound [ t]. If the second phoneme ( [r ] ) is detected, all the words from the cohort will be deleted that do not begin with the phoneme [tr ]. If the cohort, for example, recognition of the [t ] of the words tree ( tree), trespass ( trespass ) time, ( time), train ( train ) tress ( Locke ), then with recognizing the second Loud the time of the cohort removed because it does not begin with the sound sequence [tr ]. If the third volume detected ( [ ɛ ] ), the cohort will trespass on the entries, tress and shortened workout ( tree is phonetically realized as [ tri ː ] ). Only with the recognition of the fifth sound, [ p], is only a word in the cohort is (namely, trespass ), and the chain of sounds than the word trespass clearly recognized.

Explanation of context effects

In the cohort model, the faster detection of bound by the context lexemes by adopting the parallel activation is explained. This assumption is that all members of a cohort likewise be activated, that is, the listener accesses mentally to all members of the cohort. This activation is then withdrawn when the lexeme no longer matches the given input if the lexeme is thus deleted due to a different phoneme from the cohort.

Since the activation of individual lexemes all the information of the lexeme, ie morpho - syntactic, phonological as well as semantic, are always activated as well, the handset also uses Related entries in the mental lexicon. These are other entries in the lexicon that have the same or sufficiently similar properties.

The results of the above experiments can be explained by the cohort model as follows: When understanding of the first word of this and all relatierten to him, words are capitalized. Upon detection of the following word, the activated information is still available. Will now be an indirectly activated lexemes included in the cohort of the following word, the detection of the second word takes place in a shorter time compared to unrelatierten and thus not activated words.

Another consequence of the assumption about the parallel activation is that the discrepancy between uniqueness point and recognition point can be explained directly.

Criticism

Very soon after the publication of the cohort model were known issues that could not be solved without further assumptions the model. These include the frequency effect as well as dealing with a defective input.

The frequency effect states in the narrow sense that the listener a word he used frequently, quickly recognizes as another, which he used tend to often. With faulty input is meant that in spoken language seldom comes the word as a whole in the listener. Static or background noise maim in many cases part of the acoustic information, which is transmitted from the speaker to the listener. Nevertheless, the receiver, in most cases, be able to understand the spoken language.

This criticism came largely from James L. McClelland and Jeffrey L. Elman and subsequently led to the development of the TRACE model, an alternative to the cohort model, but which is the same on certain essential elements of recourse

To address the problems, advanced Marslen -Wilson after a series of experiments, his model, which is known in the current literature as Cohort II.

Cohort II

Soon after the publication of criticism of his model, Marslen -Wilson led a number of further experiments by to check these effects, and changed his model then the results accordingly. The experiments, the results thereof and the modifications to the model explained Marslen -Wilson in an article from 1987 ( see References ).

Frequency effect

To test the effectiveness of the frequency effect, let Marslen - Wilson in the mid-1980s re-run experiments. In these it was so-called lexical decision experiments with visual target words. In these experiments, the subjects given the task of words that they get displayed on a screen, to decide by pressing a button whether they are words of their language or not. In addition, the subjects were presented through headphones different words (called Destruktorwörter ). Was varied in these experiments, the timing at which the word on the screen in relation to the heard word was displayed.

The adjacent table illustrates the results of this experiment. The subjects were presented acoustically, for example, the words captain (English " captain " ) or captive (English " prisoner "). The corresponding lexialen decision to visually presented words were ship (English " ship" ) or guard (English here: " guards "). The time at which the word was presented on the screen could be (shown in the table by the " ‣ " ) immediately prior to the T or at the end of the acoustically presented word.

The word captain is assumed to be more frequent as a captive, which means that captain in the average vocabulary of a native English speaker tends to occur more frequently than captive. As the table shows, the subjects recognized the word ship faster than the word guard when it became the early stage (ie, immediately before the included / t / ) presented. This is explained by the fact that the word captain in understanding the first three phonemes (cap ) is activated and thus the word may prime ship, while the other possible word ( captive ) was not or weakly active, whereby a priming effect for guard delayed to the detection lasted longer. However, exposing the words to be recognized presents late, this effect is lost. Marslen -Wilson concluded that the frequency effect indeed true, but acts only early and will be overwritten at a later time through general context effects and loses its effectiveness.

In order to explain the frequency effect with his model, let Marslen - Wilson in the second version of the model, the all-or -nothing assumption fall and replaced it with a goodness - of-fit approach (English, mutatis mutandis: what works best fits ). Where, on the former members of a cohort yet all equally activated and deactivated, it is assumed in the second version of the model that certain entries in this glossary a larger activation potential than others. Frequente entries are within a cohort is more strongly activated than less -frequency, which is intended to explain the earlier activation of the frequenteren words.

Defective input

Various experiments have shown that subjects are able to recognize words when certain parts of the words were garbled by noise, such as noise or by loading a different sound. If you play, for example, the word embedded the subjects universal in a set before, replacing s by noise and asks the subjects where the error was, so it falls the most difficult to locate the fault exactly, let alone that she did a mistake notice. Allowed in place of s, however, a silent gap, the subjects recognize the error in nearly all cases correctly.

The original version of the cohort model works on phoneme level. The linguistic units which are recognized serial, here are the sounds of the word to be recognized. To explain the observation of understanding of utterances in spite of potential noise, was dropped this assumption and changed the model so that it is now working with distinctive features. In theory sounds can be decomposed into several distinctive features, so this service is the sound / t / the features [- voiced, CORONAL, - sonorantisch, -nasal ] and so on.

In the second version of the model the cohorts no longer be opened after recognizing certain sounds, but when an undetermined amount is given to the overlapping phonological features. This entails that the entries in the mental lexicon phonologically not ausspezifiziert, and are therefore stored as a string of phonemes, but in the form of chains of distinctive features. If now a part of these features, dubbed by noise, the entries remain in the cohort, the remaining features coincide.

Criticism

As an important problem of both versions of the cohort model, the definition of words is called. The model provides for no mechanism can be seen in a coherent text, the beginning and end of a word. Nevertheless, it provides a very robust mechanism to explain the recognition of isolated words.

Recent Developments

The cohort model is valid despite its weaknesses today as a standard model of the auditory word recognition, which has been integrated into its fundamentals in many later models.

The successive exclusion of inappropriate elements of a candidate set can also be found in the optimality theory again, a formal model of the grammar of human languages.

Also in computational linguistics basic features of the model were integrated. The truncation of the research in database systems makes use of the basic mechanism of the reduction of a relevant result set by segment- wise results exclude potential advantage. It thus forms a direct application of the cohort model in computational linguistics.