|Cognitive psychologists have discovered a body of repeatable effects when
examining people's abilities to recognise words. A typical experiment involves
showing a person a series of words and asking them to respond as quickly as
possible by pressing one button if they recognise it as a proper word and another
button if they do not recognise it as a proper word. The time taken to respond
is a key parameter. A typical finding is that common words are recognised more
quickly than uncommon ones (the word frequency effect). Often, priming words
are used in the experiments (i.e. words shown before the target word). It has
been found that a priming word can accelerate a response (the semantic priming
effect), e.g. showing the word "bread" before the word "butter"
will speed-up recognition of "butter".
This essay compares two models of word recognition that attempt to account for the basic findings from such experiments. The first part of the essay described the way in which the Logogen model accounts for basic word recognition findings. The second part contrasts the Logogen model's effectiveness at handling such data with the Verification model.
According to the Logogen model, the word frequency effect is explained by logogens having different thresholds, such that "logogens corresponding to words of high frequency in the language have lower thresholds" (Morton, 1969). Hence, high frequency (i.e. common) words require less perceptual information to raise their activation to threshold, hence are recognised more quickly than low frequency words.
Degradation effects can be accounted for in two different ways: (1) A degraded stimulus requires a stage of 'cleaning-up' (or normalising) prior to the lexical system, which adds a fixed time to orthographically similar words irrespective of their frequency (Besner & McCann, 1987); or (2) "The sensitivity of the logogen falls in line with any reduction in stimulus quality" (Mitchell, 1982) i.e. the rate of increase of a logogen's count (aka activation) is slower with a degraded stimulus, because the rate of feature extraction is slowed. The latter proposal would, therefore, predict an interaction effect between stimulus quality and frequency, whereas the former proposal predicts an additive relationship. We shall see later the empirical support for, and against, these predications. Non-word effects are robust findings from LDT experiments which must be accounted for in any comprehensive word recognition model. However, it is worth noting that Morton only intended the Logogen model to be applied to tachistoscope recognition and not to LDT experiments, hence the basic model lacks any mechanism for explaining non-word effects (since non-words do not have logogens). Coltheart et al (1977), therefore, added to the basic model in order to account for certain non-word effects.
There are three basic non-word effects (Taft, 1991):
(1) The non-word legality effect (e.g. CRVSS gets a much quicker 'no' response than does CRUSS). To account for this effect, Coltheart et al proposed that there is a variable deadline which is adjusted during processing, depending on the overall level of excitation in the logogens; the deadline being raised if excitation is rising rapidly (since the stimulus is probably a word), but left unchanged if the system is fairly quiescent. An illegal non-word causes little excitation, which results in a low deadline, hence a quick 'no' response
(2) The word similarity effect (e.g. TRUAN gets a quicker 'no' response than does TRIAN; the latter being very orthographically similar to TRAIN) is accounted for in the same way. That is, the more a non-word looks like a real word, the higher the excitation in the logogen system, and the longer will be the deadline. This effect also results in more LDT errors (i.e. TRIAN gets more 'yes' responses than does TRUAN). This is simply accounted for by a false firing: i.e. a word like TRIAN is likely to raise the activation of the TRAIN logogen to threshold, since it has very many of the features that this logogen is sensitive to
(3) The lexical status effect (e.g. BLINK gets a quicker 'yes' response than FLINK gets a 'no' response) is accounted for by the Coltheart's idea that a legal non-word causes moderate excitation in the system, hence a raised deadline; this deadline being longer than the time it takes for a real word to be accessed (which would have to be the case to avoid identifying lots of real words as non-words).
The final basic word recognition finding to be examined is the word similarity effect as applied to real words (e.g. BEAGLE gets a quicker 'yes' response than does MINUET; the latter being very orthographically similar to MINUTE). This is explained by assuming that increases in overall system excitation cause increases in all logogen thresholds. Hence, a word like MINUET which strongly excites, not only its own logogen, but also the MINUTE logogen, causes more overall excitation than BEAGLE (which only strongly excites its own logogen). Thus, the MINUET logogen has a raised threshold, compared to the BEAGLE logogen, hence a slower recognition time.
The effectiveness of the Logogen model to account for the basic word findings is most stringently tested when word frequency, context and degradation vary together. Since frequency and context (i.e. semantic priming) are both accounted for by altering logogen thresholds, these effects should show an additive relationship. Although an additive relationship was found by Schuberth & Eimas (1977), subsequent investigators, using improved stimulus materials and modified procedures, have found that frequency and context interact (Becker, 1979; Borowsky & Besner, 1993) such that the effect of context is larger for low-frequency words than for high-frequency words. This poses a major problem for the basic Logogen model.
Two predictions can be made for the joint effect of both degradation and frequency, and context and frequency, depending on which of the degradation mechanisms is assumed - though both mechanisms, being prior to the lexical system, should result in degradation having the same pattern of interaction with both frequency and context. The two predictions are: (1) assuming a fixed time is added due to a 'clean-up' stage, then degradation and frequency should be additive, and degradation and context should be additive; (2) assuming the rate of feature extraction slows with degradation, then degradation should interact with frequency, and degradation should interact with context.
Unfortunately, for this model, neither of these predictions is fully true since degradation and frequency show an additive relationship (Becker & Killion, 1977; Borowsky & Besner, 1993), and degradation and context show an interactive relationship (Schvaneveldt et al, 1976; Becker & Killion, 1977). This a second major problem for the Logogen model.
In contrast to the Logogen's difficulty in accounting for this evidence, is the Verification model any more effective? The model considered here will be the Activation-Verification model based on Becker's work (1976, 1979, 1985) and extended by Paap and colleagues (1982, 1987).
First, let me very briefly outline how this model accounts for word frequency, context and degradation effects, before looking at interaction effects. Frequency effects are accounted for by the ordering of the sensory set (aka candidate set), so that high-frequency candidates are verified early and low-frequency candidates verified late. Context (semantic priming) is accounted for by a semantic set which is created by a Semantic System and is verified against the target's visual representation before the sensory set is verified. If the target is not semantically congruent, then the semantic set will be fully searched before the sensory set is searched; hence, primed words are matched more quickly. Degradation is assumed to exert its effect on two different processes: (1) slowing of the system which creates the sensory set; (2) slowing of the construction of the visual representation against which the verification task takes place.
What then are the interaction effects arising from this model? Since the semantic set is not frequency ordered, both high-frequency and low-frequency words will take about the same time to be matched if in the semantic set. Hence, context and frequency interact, since low-frequency words are affected more by context than high-frequency words. This interaction is consistent with the empirical findings mentioned earlier. Degradation increases the time taken to generate the sensory set and the visual representation, but has no effect on the duration of the verification process. Therefore, frequency and stimulus quality should be additive, since both high-frequency and low-frequency words are delayed by the same amount. Again, this is consistent with the evidence discussed earlier.
It is a central assumption that the effect of degradation on the time taken to create the sensory set is substantially larger that the effect on the time taken to create the visual representation. Thus, for a degraded, non-contextual target, the sensory set will not be completely formed by the time the semantic set has been fully searched. The effect of degradation on a contextual target will just be the extra time taken to create the visual representation. Hence, degradation and context interact, which is consistent with the empirical findings described earlier.
Although this model appears to effectively explain the basic findings and joint effects, it is not without its problems. Let me very briefly outline five of the problems.
(1) The assumption that the time taken to create a sensory set from a degraded stimulus is greater than the time to create the visual representation plus the time taken to search the semantic set, is "apparently arbitrary" (Besner & Smith, 1992) and leads the counterintuitive prediction that for very large semantic sets (such as for fruit) context and degradation may have an additive relationship (since this assumption may not hold).
(2) The model cannot account for the finding that repetition priming interacts with degradation (Norris, 1984), a finding which the Logogen model can account for.
(3) The model assumes that the verification order of the semantic set is due to associative strength with no word frequency effects. In fact, a partial frequency ordering of the semantic set may be required to account for the fact that frequency does have some impact on processing of primed targets (c.f. Becker, 1979).
(4) The model is based on the fundamental assumption that creation of the sensory set uses only "primitive" features, whereas verification uses a more structurally oriented "relational" matching process. There is little supporting evidence for such a distinction, as Becker himself acknowledges (Becker, 1985).
(5) Although this model is relatively effective at covering the basic word recognition findings, it has problems explaining: (i) neighbourhood size effects; (ii) the standard finding of "facilitation without inhibition" seen at short SOAs (Neely, 1976); and (iii) the particular word similarity effect whereby a possible delay in response to a high frequency word arises from its similarity to a low frequency word.
In conclusion, it is apparent that both the basic Logogen model and the Verification model account for the basic word recognition findings quite effectively, although the Verification model is much better at explaining joint effects of word frequency, degradation and semantic priming. However both models are in need of modification if they are to provide comprehensive and effective accounts of all word recognition findings.
Becker, C. A. (1976). Allocation of attention during visual word recognition. Journal of Experimental Psychology: Human Perception & Performance, 2, 556-566.
Becker, C. A. (1979). Semantic context and word frequency in visual word recognition. Journal of Experimental Psychology: Human Perception & Performance, 5(2), 252-259.
Becker, C. A. & Killion, T. H. (1977). Interaction of visual and cognitive effects in word recognition. Journal of Experimental Psychology: Human Perception & Performance, 3, 389-407.
Besner, D & Smith, M. C. (1992). Models of visual word recognition: when obscuring the stimulus yields a clearer view. Journal of Experimental Psychology: Learning, Memory & Cognition, 18(3), 468-482.
Borowsky, R. & Besner, D (1993). Visual word recognition: a multistage activation model. Journal of Experimental Psychology: Learning, Memory & Cognition, 19(4), 813-840.
Coltheart, M., Davelaar, E., Jonasson, J. T. & Besner, D. (1977). Access to the internal lexicon. In Attention and performance VI. Dornic, S (ed). New Jersey: Earlbaum.
Garnham, A., (1985). Psyholinguistics:central topics. London: Routledge.
Norris, J. (1984). The effects of frequency, repetition and stimulus quality in visual word recognition. Quarterly Journal of Experimental Psychology, 36A, 507-518.
Mitchell, D. C. (1982). The process of reading: a cognitive analysis. Chichester: John Wiley.
Morton, J. (1969). Interaction of information in word recognition. Psychological Review, 76, 165-178.
Taft, M. (1991). Reading and the mental lexicon. UK: Earlbaum.
Schuberth, R. E. & Eimas, P. D. (1977). Frequency of context on the classification of words and non-words. Journal of Experimental Psychology: Human Perception & Performance, 3, 27-36.
Schvaneveldt, R. W., Meyer, D. E. & Becker, C. A. (1976). Lexical ambiguity, semantic context, and visual word recognition.Journal of Experimental Psychology: Human Perception & Performance, 2, 243-256.