Models of visual word recognition
المؤلف:
Paul Warren
المصدر:
Introducing Psycholinguistics
الجزء والصفحة:
P144
2025-11-08
43
Models of visual word recognition
The various findings from research on visual word recognition have led to a number of models of the process of recognising words. A selection of three dominant models will be presented in this section, in order to illustrate the different mechanisms that have been suggested for accounting for some of the basic effects outlined above.
Morton’s Logogen model
In the ogogen model (Morton, 1969), words are represented by logogens, recognition units that are activated on the basis of different types of input information. One way of conceptualising this is to see the logogens as containers, into which the input information is poured see Figure 9.1. Because this information can include phonological input as well as ortho graphic, the Logogen model is also a model of spoken word recognition.
The logogens can also be activated by appropriate contextual input information. Word recognition occurs when the activation level of a recognition unit climbs above some threshold. The logogen-as-container analogy in Figure 9.1 shows how different words may have different thresholds the dotted lines, so that the same input information the shaded content of the units brings one word over its threshold before the others. Such threshold differences might result from the frequency with which words are encountered, with repeated exposure to a word causing a lowering of the logogen threshold.
Since there is no interaction between the recognition units, the Logogen model has no obvious mechanism for accounting for the effect of neigh bourhood size.
In some respects the Logogen model is similar to the Cohort model. It is clearly a parallel model, with more than one word candidate activated at any one time. It also proposes that the initial activation of candidates is a result of bottom-up input i.e. in this case orthographic or phonological input, with context effects only coming into play once logogens have been activated.
Forster’s Search model
The Search model is a serial search model (Forster, 1976; Murray & Forster, 2004). It differs in this respect from both the Cohort model discussed in Chapter 8 and the Logogen model presented above, which are parallel models. In serial search models, the checking of the input against a candidate set of words is done one word at a time. Frequency effects are easily accounted for in a serial model by the ordering of words in the candidate set being checked. High-frequency words are checked first, and so high frequency words are recognised more quickly than low-frequency ones.
The candidate set that is being checked is determined in the Search model by the type of input, which depends on the task at hand. The model has access files for orthographic, phonological and syntactic/semantic input see Figure 9.2. There are multiple access files of each type, organised according to relevant factors, so that for instance words with the same sequences of beginning letters will be listed in the same bin. Only one access file of each type is shown in the figure. The access files or bins’ have links to a master file, i.e. a complete listing of words in the mental lexicon. This arrangement is somewhat similar to the organisation of a library, with the access files being a type of catalogue, giving a code for each word just as the library catalogue gives a code for each book. These codes allow the processor to find the word in the master file, in the way that book codes allow the user to locate the book in the relevant stack or shelf.
Although different in other respects, the Search model is similar to the Logogen model in that it cannot easily account for neighbourhood effects. In this case this is because the serial search process means that words are considered one at a time and are not compared all at once to the set of their potential competitors. In particular, there is difficulty accounting for the finding noted above of facilitatory effects of neighbourhood size on lexical access. As noted by (Andrews 1992), an increase in neighbour hood size would be expected to result in slower responses because of the increased number of similar words to be checked. At best, there would be no effect of neighbourhood size, if the neighbours were not sufficiently similar to the target word to influence the search for it.
An Interactive Activation model
The models outlined above assume that the visual recognition of words follows a linear process whereby recognition elements such as letters are detected in the input and mapped onto representations for words in the lexical system. In Chapter 7, however, we saw that the perception of both the spoken and the visual input is affected by linguistic information. This included the Ganong effect Ganong, 1980, which we saw resulted in a shift in perceptual boundaries between two categories (e.g. /d/ and /t/ phonemes), depending on which category resulted in a real word (e.g. desk vs tesk and dask vs task). In an attempt to characterise such findings, computer based implementations of word recognition models have been constructed which allow the flow of activation not only from the recognition elements at the input stage to the lexical system but also from the words in the lexical system back to the recognition elements. These are known as Interactive Activation IA models (McClelland & Rumelhart, 1981 Rumelhart & McClelland, 1982; Seidenberg & McClelland, 1989).
The workings of such a model cf. the TRACE model are illustrated in Figure 9.3, where the three levels of units are letter features, letters and words. The connecting lines between the units in the model are of two types. Some links, ending in arrows in the figure, are excitatory. That is, if one unit receives activation, then it will pass this activation on to units at the other end of such excitatory links, following the direction of the arrow. The other type of link is inhibitory, and is shown by lines ending in dots. If two units are connected by such a link, then activation of one will cause inhibition, i.e. reduction of activation, in the other.
As input activation is received at the level of letter features, it is passed up through the model to units for letters that contain those features. At the same time, inhibitory connections ensure that letters which do not contain those features, i.e. which are in this respect incompatible with the input, have their activation levels lowered, making them less likely for recognition. Similar activation and inhibition links are found between the letter units and the word units. Note also that in this illustration links between units at the same level are inhibitory in nature, so that if one letter feature, letter or word unit is activated, then it suppresses the activation of competing units. In addition, there are links that allow activation or inhibition effects to flow down through the model from words to
letters and from letters to letter features. This means for example that as a word receives activation on the basis of its initial letters, so this word in turn feeds activation down to each of its component letters. This raises the activation level of letters not yet encountered, with the effect that these letters are more readily recognised. It is easy to see how such a system would predict the Ganong effect – letters or phonemes are more easily recognised if they receive activation from the lexical level.
Although this is not shown in the illustration, IA models can be extended to include units representing the semantic and syntactic properties of words, so that words are more readily recognised if they fit the context in which they are found. Note also that some IA models, such as that proposed by Seidenberg and McClelland 1989 are distributed’ models, and do not have representations corresponding to linguistic entities such as words. Instead, letters, words, etc. emerge as patterns of activation across elementary processing units in the interactive network. Additional refinements of this class of model include variations in the strengths of the connections between units, to better reflect specific patterns of competition and support. Individual units can also have different resting levels of activation, capturing the frequency effects found in behavioural research. Using such manipulations, researchers have run computer implementations of this type of model to test their hypotheses about a range of possible effects in word recognition and other aspects of comprehension.
الاكثر قراءة في Linguistics fields
اخر الاخبار
اخبار العتبة العباسية المقدسة
