The empirical work on the environment of memory that has been conducted has revealed a striking correspondence between the structure of the environment and the pattern of performance in human subjects. Anderson & Schooler (1991) have shown that:
The word frequency effect is one of the clearest examples in established research of the environment impacting upon memory performance and, hence, is a natural place to begin the analysis. The null list strength effect, which has received a significant amount of attention in recent years (Murnane & Shiffrin, 1991a; Ratcliff et al., 1990; Yonelinas, Hockley, & Murdock, 1992; Shiffrin et al., 1990; Murdock & Kahana, 1992; Chappell & Humphreys, 1994; Chappell, 1993; Heathcote, 1993), is not so obviously contingent on the environment but provides a strong constraint on any model of memory and is a test of the generality of environmental analysis.
When undertaking a study of the environment there are at least two possible approaches. The first is to build a model of the environment and then demonstrate its validity. Essentially, this is the approach adopted by Anderson (1990). Such a methodology has lead to criticism from commentators such as Gigerenzer (1991) who argue that Anderson "seems to have started with Bayes' theorem as a model of rationality and to have assumed that the structural assumptions underlying Bayes' theorem specify the structure of the environment as well."(p 496, Gigerenzer, 1991). However, the environment can be studied independently of any Bayesian assumptions. Experimental phenomena could be considered on a case by case basis by identifying the everyday tasks in which they are involved and collecting the appropriate statistics. This is closer to the approach used by Anderson & Schooler (1991) and will be adopted here.
One factor not considered explicitly in Anderson & Scholler's (1991) work is the impact of context and contextual shifts. The retention intervals are measured in days or utterences but whether previous occurrences might be thought to be coming from the same or different contexts is not explored. However, in recognition tests, for instance, there seems to be an effect of the number of items within a context above and beyond that which can be explained by lag alone (Ratcliff, Clark & Shiffrin, 1990). In the environmental analyses conducted in this paper, context will be included explicitly.
In the next section, corpi from the Connectionist mailing list and the Sydney Morning Herald (Dennis, 1995) are analyzed with respect to the word frequency effect. In the following section, a corpus from the Minnesota Daily (campus newspaper) as well the the Sydney Morning Herald corpus are analyzed with respect to the item strength and null list strength effects. In each section, the first part reviews the experimental data and current accounts and the next determines the environmental statistics relevant to the phenomena.
Word Frequency in Recognition
The frequency with which a word occurs in general usage is an
environmental variable that has been shown to have an effect on
recognition performance. Many studies have demonstrated that low
frequency words are recognised more accurately than high frequency
words (e.g. Gorman, 1961; Schulman & Lovelace, 1970). On the surface,
this finding seems to contradict an account based on optimisation to
the environment. If high frequency words are seen more often, an
optimising mechanism should be adapted to process them more
accurately than low frequency words. The word frequency effect, then,
represents an important test of the environmental approach.
What makes the low frequency advantage in recognition surprising is that on a number of related phenomena subjects perform better on high frequency words. If the subject is asked to determine if a given letter string is a word (i.e. lexical decision), they do so more quickly on high frequency words (Howes & Solomon, 1951). In addition, subjects recall high frequency words more accurately than low frequency words (Gregg, 1976; Deese, 1960; Hall, 1954; Sumby, 1963), although, the pattern is not so clear in mixed lists (see Gillund & Shiffrin, 1984, for a review). Even within recognition, the word frequency effect reverses for very low frequency words with which subjects are unlikely to be familiar (Mandler, Goodman, & Wilkes-Gibbs, 1982; Rao, 1983; Schulman, 1976; Zechmeister, Curt, & Sebastian, 1978). Many of the earlier accounts of the word frequency effect relied on variables correlated with frequency such as number of associations or concreteness (Allen & Garton, 1968; McCormack & Swenson, 1972; Schulman, 1967), degree of proactive inhibition (Gorman, 1961; McCormack & Swenson, 1972; Shepard, 1967), structural or orthographic distinction (McCormack & Swenson, 1972; Schulman, 1967; Zechmeister, 1969, 1972) and pre-experimental recency (Kinsbourne & George, 1974). Subsequent studies discounted the importance of concreteness (Gorman, 1961) and the number of associates (Kinsbourne & George, 1974), and it has been argued that while the manipulation of these correlated variables does alter performance, frequency has an effect that is independent of any of the variables considered thus far (Underwood & Freund, 1970; Underwood, 1972; Kinsbourne & George, 1974).
The first mathematical model to be applied to the word frequency effect in recognition was based on the assumption that the recognition decision was made by considering the relative increment in familiarity (Mandler, 1980). If d is the absolute increment caused by studying a word and, H and L are the high and low frequency base familiarity rates respectively, then low frequency words will be better recognised since d/(d+L) > d/(d+H) . Furthermore, since high frequency words are familiar, they may not be attended to as thoroughly as low frequency words. Consequently, d may be smaller for high frequency words, thus accentuating the difference (Mandler, 1980).
As mentioned above, one of the striking features of the empirical data is that word frequency alters recognition and recall performance in opposite directions. The SAM model (Gillund & Shiffrin, 1984) not only accounted for the low frequency advantage in recognition, but also the high frequency advantage in recall and the diminishing of this effect in mixed lists (i.e. lists that contain high and low frequency words). It was assumed that:
All of these accounts suggest that additional learning or familiarity should make a given set of words behave more like high frequency words. Allen and Garton (1968), however, conducted a study in which they tested both physics students and arts students on common and physics related (i.e. rare) terms. Figure 1 shows the operating characteristics they observed. The first point to note is that the comparatively rarer physics words were better recognised than the common words by both student populations. Hence, the usual word frequency effect was found. However, the performance of the physics students on the physics terms was much better than the arts students. Zechmeister et al. (1978), when interpreting these results, suggested that the inferior performance of the arts students indicated that the physics words were acting like very low frequency or nonwords for the arts subjects. However, since the arts students performed significantly better on the physics words than the common words it would seem that they were capable of identifying the words, and, hence, it seems unlikely that the arts students were treating them as very low frequency words. Presumably, the physics students had been exposed to these words with greater frequency than the arts students. To the physics students, then, the physics terms should have appeared to be high frequency and, in keeping with frequency based explanations, they should have performed worse on these items.
The same logic applies to most variables that correlate with frequency. The physics students would be expected to have formed more associations with the physics words and, hence, association explanations predict a decrement in performance (Allen & Garton, 1968). Likewise, if proactive interference from preexperimental words were the explanation, the physics students who are likely to have been exposed to many similar physics words would be more likely to experience such interference and should have performed worse. The orthographic or structural distinctiveness of the physics words would, if anything, be judged to be weaker by the physics students, since the physics words were familiar. Finally, the pre-experimental recency could not account for the performance of the physics students since it is more likely that they would have seen the words more recently than the arts students. The effect of subject knowledge provides strong evidence that none of the variables examined so far in the literature is the primary cause of the word frequency effect in recognition.
Allen and Garton's (1968) result seems even more striking when compared to a similar experiment in which lexical decision performance was observed for engineers, nurses and law students on a variety of domain specific terms (Gardner, Rothkopf, Lapan, & Lafferty, 1987). Just as was observed in the recognition experiment, performance was best for subjects who were familiar with the materials. In lexical decision, such a result is not so surprising since high frequency words are identified as words more quickly than low frequency words. These phenomena establish, however, that training seems to uniformly improve performance. Why, then, is the effect of frequency in recognition the opposite to that in lexical decision?
A major difference between the recognition and lexical decision tasks is that the recognition as assessed in the laboratory is episodic, whereas lexical decision is not. If the memory system is optimized to the environment it may be sensitive to contextual boundaries producing a difference between recognition and lexical decision. Consequently, in the following environmental analysis, statistics relevant to recognition are collected within an episode or context.
"The Brown corpus, which was compiled with the view of making it broadly representative of current edited American English, contains selections from five hundred samples belonging to fifteen different genres of writing. The genres range from newspaper reportage to technical writing, and from philosophical essays to various kinds of fiction."(p 1, Francis & Kucera, 1982)
In contrast, recognition experiments usually occur within a single context. The recognition question is episodic. Subjects are required to recognise an item from a specified context.
Furthermore, Francis and Kucera (1982) note that the distribution of words in different genres is not even. They point out that unlike high frequency words, low frequency words tend to occur in a smaller number of genres. That is, they seem to be context specific, a notion that has some intuitive appeal. For instance, in a message posted to the connectionist mailing list the word "RAAM"[1] was used six times within an article of some 150 words. The article was about "RAAMs" and, hence, it is not unusual that the word was repeated several times. The word "RAAM", however, is clearly a very low frequency word in English and certainly does not occur in the Kucera and Francis (1967) word frequency inventory.
However, if low frequency words are more likely to be context specific, they may be repeated frequently within the contexts to which they pertain, and hence, the memory system would be optimised to recognise them when they recur within these contexts. The hypothesis to be tested in this analysis, then, is that low frequency words recur more often within a context than high frequency words.
As in the analysis of Anderson and Schooler (1991, i.e. the New York Times and email analyses), reading is assumed to require retrieval of the words. More specifically, it is presumed here that as people read text they make an episodic recognition decision for each word and that reading is one of the dominant real world tasks to which episodic recognition performance would be optimised. However, this assumption requires some justification.
Episodic Recognition in Reading: In the case of proper nouns (e.g. John, Mary), an episodic recognition decision must be made. If a person reads a second occurrence of "John", they must recognise that the name has already occurred in this context to realize that the referent is probably the same in each case. A similar argument can be made for pronouns, although, there must also be an episodic cued recall component (i.e. determining to whom the pronoun refers).
Furthermore, the processing of nouns in general would seem to include episodic recognition as a component. For instance, consider the sentence "Shut the door.". If a door has been mentioned in the current context then it is likely that it is the door to which the sentence refers. Realizing that "door" has already occurred is an episodic recognition decision.
While the case is clearest for nouns, other word classes would require episodic recognition also. For instance, if one talks about something that is "green", realizing that green things have already been mentioned in the current episode is important. Hence, the processing of adjectives would sometimes require an episodic recognition decision. Similarly, verbs may also require an episodic recognition decision. If the word "whistle" recurs there will often be a link between the occurrences.
The major exceptions are the functor words such as "of", "the" or "for". These words recur often and their episodic identity seems to be of little consequence. There is some indication that function words are not processed in a serial fashion and it may be that no episodic recognition decision is made in this case. Further, function words are not usually included in laboratory paradigms. Hence, these words were not included in the analysis (c.f. Anderson & Schooler, 1991).
It might be argued, however, that episodic recognition is not involved in the reading of words because the words from the current context would be transferred to some shorter term data structure[2]. Rather than use a contextual cue to retrieve words from episodic memory such a system would search the short term structure to determine if a word had occurred in the current context.
However, Brannelly, Tehan, and Humphreys (1989) present evidence which suggests that even in a situation in which words might be expected to be in a primary memory structure there would seem to be a retrieval component which is subject to proactive interference and, hence, would seem to involve episodic (secondary) memory. In their second experiment, Brannelly et al. (1989) presented subjects with memory sets of either two or four items. After a filled 14 second delay, subjects were presented with a probe which they were required to identify as having been in the memory set or not. A second probe was then presented and again the subjects were required to determine if it had been present in the memory set. In addition, proactive interference was manipulated by choosing items from the same taxonomic category on three consecutive trials. The assumption was that subsequent to the presentation of the first probe the memory set would be retrieved from secondary memory into primary memory and scanned to establish the presence of the probe item. When the second probe was presented one would expect that the memory set would still be present in primary memory and, hence, should not be affected by proactive interference which is usually a feature of retrieval from secondary memory. However, there was a significant amount of proactive interference. Hence, even if the words from the current context are present in some short term data structure it is likely that there will also be retrieval from secondary memory. It could be argued, however, that the short term data structure would be longer term than the primary memory which was was implicated in the memory probe tasks used by Brannelly et al. (1989) and may be subject to proactive interference. The question becomes "Can such a memory system be distinguished empirically from episodic memory?".
Humphreys and Burt (1994) argue on the basis of data from Jacoby (1983, experiment two), that episodic information affects repetition priming even when context is not reinstated, but is implicit because the test occurs immediately following the study opportunity. In Jacoby's second experiment, subjects were presented with a study list and then with a surprise perceptual identification task. The saliency of the study items was manipulated by administering the test either immediately in the high saliency condition, or after 24 hours in the low saliency condition. In both cases the test list consisted of 10 words from the study list and 90 new words. The subjects were not informed about the relationship between the study and test lists. The perceptual identification of old words was facilitated in the high saliency condition relative to the low saliency condition. Jacoby (1983) also found inhibition of the new words in the high saliency condition relative to the low saliency condition. As Humphreys and Burt (1994) point out, such a finding can be explained by assuming that an episodic memory system is employed in both conditions and that the context cue in the high saliency is more similar to that at study than the context cue from the low saliency condition. It is also possible to explain these results using a short term data structure by supposing that the subjects were using this structure to make their decisions in the high saliency condition and were not in the low saliency condition. Under such a view, the old words would show facilitation in the high saliency case because the short term store is searched first. The inhibition of the new words in the high saliency case relative to the low saliency case would occur because of the loss during the search of the short term datastructure that occurs in the high saliency case but not in the low saliency case.
What this analysis suggests is that it may be difficult to distinguish the consequences of the short term data structure from an episodic memory system which allows the use of context cues which persist as a consequence of continuity rather than requiring explicit reinstatement.
In terms of using reading as the environmental analog of episodic recognition tasks, the distinction may not be important. If there is a short term structure, one must assume that it is under the strategic control of the subject to account for the Jacoby results. If this is the case, it would also be available to subjects in making recognition judgements in the "episodic" tasks that have been studied. Regardless of whether it is a list context that is reinstated or a short term datastructure that is identified as relevant it is still the recurrence probability that one would expect to be predictive of performance on these tasks.
The above discussion is not meant to suggest that episodic recognition is the only subtask of reading, nor that reading is the only task in which episodic recognition takes place. As noted above, reading must include cued recall in order to determine to whom or what pronouns refer, and there are most probably many other situations in which cued recall would be implicated. Other tasks, notably conversation, would also seem to require episodic recognition. Reading has been chosen, in this case, because it is a very common task for which large corpora are readily available for analysis.
To quantify the likelihood of recurrence of a word within the same context, the word density was defined as the number of occurrences of the word divided by the number of contexts in which it occurred.
As a further test a larger sample (23440636 items) was taken from the 1994 editions of the Sydney Morning Herald. There were a total of 38526 articles each of which was considered a separate context. Removal of misspellings and function words is prohibitively expensive on such a large corpus. To decrease the number of typographic errors counted only items which occurred in two different articles were considered. This restriction lead to a total of 97031 unique items. Using the larger sample allowed the word frequency data to be generated from the sample itself. The previous analysis used the Kucera and Francis (1967) statistics because there were very few observations with which to get stable estimates of frequency. Also the analysis of the larger sample focused on the frequency band between 1 and 100 occurrences per million which is the band most commonly considered in the experimental literature and which shows the most pronounced effect in the previous analysis. Figure 4 shows the scatter plot of density verses the log of the word frequency.
Again the slope of the regression line is negative (slope -0.016, intercept 1.44), F(1, 30758), p = .049, supporting the results from the first analysis that low frequency words are more likely to recur. It should be noted, however, that this result is not as clear as in the first analysis and is sensitive to the placement of the frequency boundaries at both ends. At the low end (frequencies below one per million) the sensitivity may be a sampling artifact. The maximum possible word density increases with the number of times the word occurs, this restriction being particularly problematic at the lower frequencies. If the frequency boundary is moved above 100 the mean density increases and the result becomes gradually less significant (see figure 5). This may be a consequence of the inclusion of a greater number of function words which will increase the word density but would not be included in a typical memory experiment.
The current data support the recurrence rate explanation of the word frequency effect. Since low frequency words are more likely to recur, an optimising memory system should adapt to perform more accurate episodic recognition on the low frequency words. However, by definition, the occurrence rate of a high frequency word that has not yet occurred in the current context will be greater than that of a low frequency word. Given the above analysis, there must be an interaction between occurrence history and word frequency (see figure 6).
Furthermore, if one considers the probability of recurrence at any given instant as a function of the time since the last occurrence of the word, low frequency words must decay more rapidly than high frequency words (see figure 7).
The interaction of word frequency and occurrence history is particularly interesting in the light of the work by Glanzer and Bowles (1976). They conducted forced choice experiments including conditions in which subjects were required to choose between a high frequency and a low frequency item both of which had been on a study list, and a high frequency and a low frequency item neither of which had been on the study list. They found that when both items had been seen, subjects were more likely to choose the low frequency item. When neither of the items had been seen, subjects were more likely to choose the high frequency item. If one assumes that performance is proportional to probability of recurrence, the Glanzer and Bowles' result maps directly onto figure 6. Recurrence rate also provides an explanation of the effect of word knowledge (Allen & Garton, 1968). While low frequency words have a higher recurrence rate in general, the recurrence rate of the physics words for physics students is likely to be greater still. The memory systems of the physics students would be optimised to the physics words and, therefore, it would be expected that they should perform better on these words than the arts students.
In addition, recurrence rate provides a straightforward account for why lexical decision should show the reverse effect to episodic recognition. When reading text a lexical decision would be made at each occurrence of a word. In contrast, a positive episodic recognition decision can only be made at the recurrence of a word in a context. Hence, lexical decision should favour high frequency words, which have a higher occurrence rate, while episodic recognition should favour low frequency words, which have a higher recurrence rate.
There is, however, a caveat to the above discussion. While a given subject may have learned that particular low frequency words tend to recur often, it is questionable whether they could have been exposed to a sufficiently large subset of the low frequency words to extract the recurrence information on a case by case basis. While the above reasoning would suggest that occurrence and recurrence frequency are the prime determinants of the frequency effect, other variables related to input representation (e.g. orthographic distinctiveness) or functional role (e.g. concreteness, or number of associates) may contribute to the establishment of a similarity landscape in which the representations of the low frequency words are separated from those of the high frequency words. Rather than learn the recurrence[3] information for each word individually, an optimising mechanism might determine that the class of low frequency words are more likely to recur. Nonwords and very low frequency words may not show the same performance pattern as low frequency words (Zechmeister et al., 1978) because they are not sufficiently similar to the low frequency words for which the recurrence response has been learned.
The Null List Strength Effect in Recognition
It is perhaps not surprising that environmental analysis proved useful
in the explanation of the word frequency effect. Word frequency is an
environmental variable. But, can environmental analysis shed light on
phenomena that are not so obviously related to the statistics of the
environment? To explore this question this section examines the null
list strength effect in recognition (Ratcliff et al., 1990; Shiffrin et
al., 1990).
The presence of repeated items in a list to be recognised does not affect performance on non-repeated items. This phenomenon, know as the null list strength effect, has undergone substantial research both empirical (Murnane & Shiffrin, 1991a; Ratcliff et al., 1990) and model theoretic (Shiffrin et al., 1990; Murdock & Kahana, 1992; Chappell, 1993; Chappell & Humphreys, 1994; Heathcote, 1993) in recent years. Unlike the word frequency effect studied in the previous section, however, the list strength effect is not so obviously related to the statistics of the environment. The purpose of this study is to investigate what impact the environment might have on the phenomenon.
Table 1: Accuracy results for nonrepeated (A) items in different list types.
| Test Type | AB vs ABB | AB vs ABC | ABB vs ABC |
|---|---|---|---|
| Recognition | = (>)* | > | >* |
| Cued Recall | = (>)* | > | >* |
| Free Recall | > | > | < |
For free recall, Tulving and Hastie (1972) demonstrated that performance on unstrengthened items dropped significantly in comparison to the AB and ABC control conditions. For recognition, however, strengthening an item does not degrade performance significantly on the non-strengthened items (known as the null list strength effect, Shiffrin et al., 1990; Ratcliff, 1990; Yonelinas et al., 1992). This result contradicted the major mathematical memory models all of which predicted that performance on A in the ABB condition should be as poor if not worse than that in the ABC condition (Ratcliff et al., 1990).
Since the initial experimental work on the null list strength effect, a number of researchers have attempted to alter their models to account for the data. Shiffrin et al. (1990) made two modifications to the SAM model. The first was to assert that when an item was strengthened, either by additional study time or additional study presentations, the same representation (or image) in memory was affected. Secondly, strengthening an item not only increased its context-to-item association, but also decreased its similarity to all other items. This process is called differentiation and is illustrated in figure 8. In the figure, familiarity is measured by a dot product. Strengthening an item has the effect of both increasing the length of the vector (i.e. context-to-item association) and making it closer to orthogonal to the other item vector. By setting the parameters appropriately, the strengthening of item one can be made to have little effect on the familiarity calculated when cuing for the weak item two. Even though the length of item one has increased, it has also become less similar and, hence, the familiarity remains constant. If the model is cued with item one the additional length of the strengthened vector can still produce the main effect of strength. Likewise, increasing the number of study vectors will lead to greater interference and a length effect will be observed.
The model by Chappell and Humphreys (1994), implemented in a connectionist architecture based on the matrix model (Pike, 1984), relies on separating the components of the system that are affected by length and strength. When an item is strengthened, context-to-central weights are updated (probabilistically). This part of the system is sensitive to length manipulations but is not affected very much by strength. The resulting representation is then subjected to an intersection operation which effectively retains only the dominant item. Then, the weights of a central autoassociator are increased and a global inhibition is learned. This network is unaffected by length but is sensitive to strength manipulations. Because only a single item is present and because memories do not interact to any great degree within this system, it is only the strength of that item that is affected. Hence, the model accounts for the null list strength effect and demonstrates the difficulty in distinguishing between distributed and nondistributed models of human memory[4].
One way of eliminating the list strength effect is to assume that pre-experimental items are also present in memory. This approach was used by Murdock and Kahana (1992) in their extension of the TODAM model. The effect of adding additional items is to swamp with noise any difference between the mean familiarity of an item that occurred with unstrengthened items and an item that occurred with strengthened items. While such a method can decrease the list strength effect to negligible levels, it also undermines the prediction of the list length effect and, hence, is not sufficient to explain the pattern of results (Chappell, 1993). This example illustrates the importance of modelling the entire pattern of recognition results. It is the building of a model that captures the list strength effect while also accounting for the item strength and list length effects that has proven difficult.
All current accounts of the null list strength effect presume that it is the mechanism of memory that is primarily responsible for the phenomenon. As yet, there has been no empirical characterisation of the environmental context of the effect. The next section aims to characterise the environmental statistics relevant to the item strength, list length and list strength effects. Some of this work has already been done by Anderson and Schooler (1991) and, hence, the analysis starts with a review of their work. Then, an environmental analysis that addresses the list strength effect and re-addresses the item strength condition is conducted.
The probability that an item will recur is linearly related to the frequency of the item (Anderson & Schooler, 1991). This relationship was found in all three sources. In the New York Times data the constant of proportionality was 1.0, in the parental speech data it was 0.76 and for the electronic mail data it was 0.9. Such a result is to be expected since words occur with a certain frequency and barring some other form of intervention, it is to be expected that they will continue to occur with the same frequency. The constants of proportionality were less than one because new words were entering the sample (Anderson & Schooler, 1991), particularly in the case of parental speech.
In addition to frequency information, Anderson and Schooler (1991) studied the effect of retention interval. In this case, they plotted the probability of an item recurring as a function of the time since its last occurrence. They found that recurrence probability showed a similar negatively accelerating curve as seen in the laboratory when interval is manipulated. Furthermore, when the log need odds were plotted against log duration, a linear relationship emerged indicating that the original curve was a power function. In the New York Times data the exponent was 0.73, in the parental speech data it was 0.77 and in the electronic mail data it was 0.83.
Anderson and Schooler (1991) provide many of the statistics required to characterise the effect of repetition in the environment. However, there are still two issues that must be resolved to give a complete picture. The first involves the independent effect of contextual frequency. Anderson and Schooler (1991) did not control for the effect of general word frequency. In the laboratory, however, it is the frequency within the study context that is manipulated. The linear relationship between occurrence frequency and recurrence rate, which Anderson and Schooler (1991) found, may have been a consequence of the general word frequency of the item. In the following study, the general word frequency has been factored out so that the effect of frequency in the current context can be assessed.
Secondly, the following study will address the question of how the recurrence probability of a word is affected by the occurrence probabilities of other words in the context - the environmental analogue of the list strength effect.
P (Wi in TestSegment | Wi in StudySegment & Wj for j!=i occur frequently in study segment)
~= P(Wi in TestSegment | Wi in StudySegment & Wj for j!=i occur infrequently in study segment) (1)
Hence, the first step in the analysis was to characterise the strength history of a given item. The number of repetitions of the word was chosen as an indication of strength rather than the amount of time the reader spent processing the word because in the environmental setting it is difficult to control or estimate the amount of time an individual spends on each item. In the two experimental lists (i.e. AB and ABB), the number of unique items is constant while the total number of items increases in the ABB list. Therefore, the general degree of repetition, Number of Items/ Number of Unique Items, is greater in the ABB list than in the AB list. The sample text can be divided into segments that have equal numbers of unique items. The number of actual items will vary, however. Some segments will have large numbers of actual items and will parallel the ABB condition. In other segments, the amount of repetition will be lower, parallelling the AB condition.
The list strength effect occurs in the experimental setting if increasing the general repetition rate decreases performance on the nonrepeated items. Correspondingly, if general repetition rate decreases the recurrence rate a list strength effect will be observed in the environmental setting.
In addition, it is important to establish that the main effect, that is, that repetition improves an item's performance, is evident in the environmental statistics. That is,
P(Wi in TestSegment | Wi occurs frequently in study segment)
> P (Wi in TestSegment | Wi occurs infrequently in study segment) (2)
If there is no effect on non-repeated items, as the experimental evidence on the list strength effect suggests, the study would provide only very weak evidence if the main effect did not prove significant. Hence, the target repetition rate, that is, the number of times the target item occurred divided by the length of the segment, was included as a variable in the analysis. In the experimental setting, increasing the target repetition rate has been found to improve performance and, hence, increasing the target repetition rate is expected to increase recurrence probability.
In the uncontrolled circumstance of general text, however, the target repetition rate may be significant purely as a consequence of general word frequency. Hence, word frequency was included as a variable in the analysis to ensure that the target repetition rate was significant independent of any contribution from general word frequency.
To summarise, if the observed experimental results are the consequence of the exposure of an optimised memory system to the environment, it would be expected that general repetition rate would have no effect of recurrence probability, but that target repetition rate would have a positive effect.
Noise components such as punctuation, function words and numerals were deleted. In order to mimic the study/test paradigm employed in typical memory experiments, a study segment and a test segment were cut from the start of each message. The study segments contained 120 unique words, although the fact that words may be repeated means that it is likely that there will be more than 120 words in total. The test segments were always 120 actual words long. If an article was not long enough to extract these segments it was discarded. Upon completion of this process 198 articles were retained. Figure 9 shows diagrammatically the division of an article.
Segments of size 120 were chosen in response to the size of the articles. Larger segments provided more accurate statistics, but required many articles to be discarded since they would not have been long enough to allow extraction of both segments. A size of 120 provided a balance between these conflicting concerns.
Calculating Recurrence Probability: The dependent variable in the analysis was the recurrence probability, a measure of how likely a given word was to recur at a position in the test segment. This statistic was calculated by taking the number of repetitions and dividing by the total number of words in the test segment (which was always 120).
Calculating General Repetition Rate: The general repetition rate was included as a measure of the repetition within the study segment. It was calculated by dividing the number of actual words by the number of unique words (which is always 120).
Calculating the Target Repetition Rate: The target repetition rate for each word was calculated by taking the number of repetitions of a word within the study segment and dividing by the total number of words in the segment.
Calculating Word Frequency: The word frequency is calculated over all articles and is simply the number of times the word occurs. For example, in the Minnesota sample the word "baseball" occurs 27 times within the 198 articles and, hence, its word frequency would be 27.
Removing Points: If a word occurred in the test set but not in the study segment it was not a possible target and was discarded for the purposes of the analysis. In addition, if the word appeared in the study segment but not in the test segment its recurrence probability was considered to be at oor and it was discarded. In the Minnesota corpus, of 17390 initial observations 3716 remained after filtering.
Table 2: Summary Statistics for the Minnesota Corpus
| Mean | Standard Deviation | |
| Recurrence Probability | 0.014 | 0.010 |
| General Repetition Rate | 1.304 | 0.122 |
| Target Repetition Rate | 0.012 | 0.010 |
| Word Frequency | 131 | 259 |
Despite this precaution, the multiple regression analysis showed no significant effect of general repetition rate on 19 of the 20 trials. The target repetition rate was significant on 17 of 20 trials and the word frequency varied the most showing significance on 14 of 20 trials. Table 3 shows the significance level of adding each of the variables given that the other two variables have already been added to the model for each of the 20 samples.
Table 3 Significance levels of adding each of the independent variables on the prediction of recurrence probability for each of 20 replications.
| N | General Repetition Rate | Target Repetition Rate | Word Frequency |
| 198 | 0.2743 | * 0.0000 | 0.7380 |
| 198 | 0.7529 | * 0.0000 | 0.5329 |
| 198 | 0.3420 | * 0.0000 | 0.1048 |
| 198 | 0.6314 | * 0.0000 | * 0.0007 |
| 197 | 0.7778 | * 0.0000 | 0.0528 |
| 196 | * 0.0295 | * 0.0000 | * 0.0031 |
| 195 | 0.7242 | * 0.0000 | * 0.0000 |
| 190 | 0.3640 | * 0.0000 | * 0.0000 |
| 189 | 0.3312 | * 0.0000 | * 0.0000 |
| 186 | 0.9261 | * 0.0004 | * 0.0000 |
| 181 | 0.2165 | * 0.0002 | * 0.0000 |
| 173 | 0.0925 | * 0.0032 | 0.5342 |
| 163 | 0.9881 | 0.0694 | * 0.0003 |
| 151 | 0.5015 | * 0.0000 | * 0.0000 |
| 138 | 0.3783 | * 0.0001 | * 0.0032 |
| 129 | 0.5412 | * 0.0190 | * 0.0000 |
| 120 | 0.4079 | 0.2661 | * 0.0000 |
| 107 | 0.7930 | 0.7128 | * 0.0012 |
| 96 | 0.8204 | * 0.0185 | 0.2114 |
| 81 | 0.1329 | * 0.0000 | * 0.0000 |
The Sydney Morning Herald corpus, which was used in the word frequency analysis in the previous section, was also analysed to determine the list strength statistics. The size of the corpus allowed a single item to be taken from each context without resampling. Of the 38526 articles in the corpus, 30564 were sufficiently long to allow the extraction of study and test segments. Table 4 provides the summary statistics of the corpus, table 5 the correlation matrix and table 6 the standardized regression coefficients, the F values and the significance levels.
Table 4: Summary Statistics for the Sydney Morning Herald Corpus
| Mean | Standard Deviation | |
| Recurrence Probability | 0.0172 | 0.0154 |
| General Repetition Rate | 1.559 | 0.1591 |
| Target Repetition Rate | 0.0147 | 0.0150 |
| Word Frequency | 929 | 14696 |
Table 5: Correlation matrix for Sydney Morning Herald corpus.
| Variable | Recurrence Probability | Target Repetition Rate | General Repetition Rate | Word Frequency |
| Recurrence Probability | 1.0000 | |||
| Target Repetition Rate | 0.6937 | 1.0000 | ||
| General Repetition Rate | 0.0072 | 0.0264 | 1.0000 | |
| Word Frequency | 0.7366 | 0.7908 | -0.0333 | 1.0000 |
Table 6: Significance tests for Sydney Morning Herald corpus.
| Predictor | Standardized Regression Coefficient | F(1, 30563) | p level |
| Target Repetition Rate | 0.2945 | 2324.15 | 0.0001 |
| General Repetition Rate | 0.0162 | 18.83 | 0.0001 |
| Word Frequency | 0.5042 | 6807.52 | 0.0001 |
The target repetition rate and word frequency are quite correlated, but target repetition rate does add significantly to the prediction when added after word frequency. The general repetition rate, however, is not strongly correlated with either target repetition rate or word frequency. While there is a significant relationship between general repetition rate and recurrence probability, the standardized regression coefficients show that this relationship is much less important than either target repetition rate or word frequency.
In addition, it was found that the recurrence probability is a monotonically increasing function of the frequency of recent use (Anderson & Schooler, 1991). However, the current study takes the Anderson and Schooler (1991) analysis one step further. A high frequency word may occur frequently in both the study and test segments by virtue of the fact that it is a high frequency word. Hence, a positive relationship between study and test set frequency may have been expected independently of whether the occurrence of a word predisposes the memory system to expect the word again within the current context. This study shows that the target repetition rate is significant even when the effect of word frequency is factored out. Hence, there is a basis for confidence in the power of the test and the finding that recent repetition is positively correlated with recurrence probability.
In conjunction with the Anderson and Schooler (1991) results, we now have a picture of the environmental context of the word frequency and list strength effects in recognition.
Discussion and Conclusions
What does the environment look like? The
environmental analysis in this chapter provides evidence in support of
the following propositions:
The word frequency result is a little more unexpected. However, since recurrence rate is most relevant to episodic recognition and occurrence rate is most relevant to lexical decision, the environmental evidence provides insight into why word frequency effects these tasks differently. Furthermore, the effect of word knowledge, which undermines previous explanations of the word frequency effect, is explained naturally by recurrence rate and an adaptive memory system.
Anderson, J. R. (1990). The Adaptive Character of Thought. Laurence Erlbaum Associates, Hillsdale, NJ.
Anderson, J. R., & Milson, R. (1989). Human memory: An adaptive perspective. Psychological Review, 96 (4), 703-719.
Anderson, J. R., & Schooler, L. J. (1991). Reflections of the environment in memory. Psychological Science, 2 (6), 396-408.
Barnes, J. M., & Underwood, B. J. (1959). Fate of first-list associations in transfer theory. Journal of Experimental Psychology, 58, 97-105.
Brannelly, S., Tehan, G., & Humphreys, M. S. (1989). Retrieval plus scanning: Does it occur?. Memory and Cognition, 17 (6), 712-722.
Brunswik, E. (1956). Perception and the Representative Design of Psychological Experiments. University of California Press, Berkeley, CA. Campbell, D. T. (1974). Evolutionary epistemology. In Schilp, P. A. (Ed.), The philosophy of Karl Popper. Open Court, La Salle, IL.
Chappell, M. (1993). Neural Network Models of Human Recognition and Cued Recall. Ph.D. thesis, The University of Queensland.
Chappell, M., & Humphreys, M. S. (1994). Analysis of a neural network with application to human memory modelling. Journal of Experimental Psychology.
Deese, J. (1960). Frequency of usage and number of words in free recall: The role of association. Psychological Reports, 7, 337-344.
Dennis, S. (1995). The Sydney Morning Herald Word Database. Noetica: Open Forum, 1(4), http://psy.uq.edu.au/CogPsych/Noetica/.
Ebbinghaus, H. (1964). Memory: A Contributon to Experimental Psychology. Dover Publications, New York, NY. Translated by H A Ruger and C E Bussenius. Original written in 1885.
Francis, W. N., & Kucera, N. (1982). Frequency analysis of English Usage: Lexicon and Grammar. Hought Miflin Company, Boston, MA.
Gardner, M. K., Rothkopf, E. Z., Lapan, R., & Lafferty, T. (1987). The word frequency effect in lexical decision: Finding a frequency-based component. Memory and Cognition, 15 (1), 24-28.
Gigerenzer, G. (1991). Does the environment have the same structure as Bayes theorem?. Behavioral and Brain Sciences, 14, 495-496. Comment on Is human cognition adaptive? by J R Anderson in same volume.
Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both recognition and recall. Psychological Review, 91 (1), 1-67.
Glanzer, M., & Adams, J. K. (1990). The mirror effect in recognition memory: Data and theory. Journal of Experimental Psychology: Learning, Memory and Cognition, 16 (1), 5-16.
Glanzer, M., Adams, J. K., & Iverson, G. (1991). Forgetting and the mirror effect in recognition memory: Concentering of underlying distributions. Journal of Experimental Psychology: Learning, Memory and Cognition, 17 (1), 81-93.
Glanzer, M., & Bowles, N. (1976). Analysis of the word-frequency effect in recognition memory. Journal of Experimental Psychology: Human Learning and Memory, 2 (1), 21-31.
Gorman, A. N. (1961). Recognition memory for names as a function of abstractness and frequency. Journal of Experimental Psychology, 61, 23-29.
Gregg, V. H. (1976). Word frequency, recognition and recall. In Brown, J. (Ed.), Recall and recognition. Wiley, London.
Hall, J. F. (1954). Learning as a function of word frequency. American Journal of Psychology, 67, 138-140.
Heathcote, A. (1993). An ART model of human recognition memory. In Leong, P., & Jabri, M. (Eds.), Proceedings of the Fourth Australian Conference on Neural Networks, pp. 212-215.
Howes, D. H., & Solomon, R. L. (1951). Visual duration threshold as a function of word probability. Journal of Experimental Psychology, 41, 401-410.
Humphreys, M. S., & Burt, J. (1994). Patterns of facilitation and inhibition in episodic and lexical access: Common functional requirements or common mechanisms?. Unpublished manuscript.
Humphreys, M. S., Wiles, J., & Bain, J. D. (1992). Direct access: Cues with separate histories. In Meyer, D. E., & Kornblum, S. (Eds.), Attention and Performance XXIV: A Silver Jubilee, pp. 489-508. MIT Press, Cambridge, MA.
Jacoby, L. L. (1983). Perceptual enhancement: Persistent effects of an experience. Journal of Experimental Psychology: Learning, Memory and Cognition, 9, 21- 38.
Jung, J. (1962). Transfer of training as a function of degree of first-list learning. Journal of Verbal Learning and Verbal Behavior, 1, 197-199.
Kinsbourne, M., & George, J. (1974). The mechanism of the word frequency effect on recognition memory. Journal of Verbal Learning and Verbal Behavior, 13, 63-69.
Kucera, H. & Francis, W. N. (1967). Computational analysis of present-day American English. Brown University Press, Providence, RI.
Mandler, G. (1980). Recognizing: The judgment of previous occurrence. Psycholog- ical Review, 87, 252-271.
McCormack, P. D., & Swenson, A. L. (1972). Recognition memory for common and rare words. Journal of Experimental Psychology, 95, 72-77.
McCullers, J. C. (1965). Type of associative interference as a factor in verbal paired-associate learning. Journal of Verbal Learning and Verbal Behavior, 4, 12-16.
Murdock, B. B., & Kahana, M. J. (1992). An analysis of the list-strength effect. Submitted to the Journal of Experimental Psychology: Learning, Memory and Cognition.
Murnane, K., & Shiffrin, R. M. (1991a). Interference and the representation of events in memory. Journal of Experimental Psychology: Learning, Memory and Cognition, 17, 855-874.
Pike, R. (1984). Comparison of convolution and matrix distributed memory systems for associative recall and recognition. Psychological Review, 91 (3), 281-293.
Rao, K. V. (1983). The word frequency effect in situational frequency estimation. Journal of Experimental Psychology: Learning, Memory and Cognition, 9, 73- 81.
Ratcliff, R., Clark, S. E., & Shiffrin, R. M. (1990). The list-strength effect: I. Data and discussion. Journal of Experimental Psychology: Learning, Memory and Cognition, 16, 163-178.
Schulman, A. I. (1967). Word length and rarity in recognition memory. Psychonomic Science, 9, 211-212.
Schulman, A. I. (1976). Memory for rare words previously rated for familiarity. Journal of Experimental Psychology: Human Learning and Memory, 2, 301- 307.
Schulman, A. I., & Lovelace, E. A. (1970). Recognition memory for words presented at slow or rapid rate. Psychonomic Science, 21, 99-100.
Shepard, R. N. (1967). Recognition memory for words, sentences and pictures. Journal of Verbal Learning and Verbal Behaviour, 6, 156-163.
Shepard, R. N. (1987). Towards a universal law of generalization for psychological systems. Science, 237, 1317-1323.
Shiffrin, R. M., Ratcliff, R., & Clark, S. E. (1990). The list-strength effect: II. Theoretical mechanisms. Journal of Experimental Psychology: Learning, Memory and Cognition, 16, 179-195.
Spence, J. T., & Lair, C. V. (1965). Associative interference in the paired-associate learning of remitted and nonremitted schizophrenics. Journal of Abnormal Psychology, 70 (2), 119-122.
Sumby, W. H. (1963). Word frequency and serial position effects. Journal of Verbal Learning and Verbal Behavior, 1, 443-450.
Thorndike, E. L. (1965). Animal Intelligence: Experimental Studies. Hafner Publishing Company, New York, NY. Original published in 1911 by Teachers College, Columbia University.
Tulving, E., & Hastie, E. (1972). Inhibition effects of intralist repetition in free recall. Journal of Experimental Psychology, 92 (3), 297-304.
Underwood, B. J. (1972). Word recognition memory and frequency information. Journal of Experimental Psychology, 94, 276-283.
Underwood, B. J., & Freund, J. J. (1970). Testing effects in the recognition of words. Journal of Verbal Learning and Verbal Behavior, 9, 117-125.
Yonelinas, A. P., Hockley, W. E., & Murdock, B. B. (1992). Tests of the list-strength effect in recognition memory. Journal of Experimental Psychology: Learning, Memory and Cognition, 18 (2), 345-355.
Zechmeister, E. B. (1969). Orthographic distinctiveness. Journal of Verbal Learning and Verbal Behavior, 8, 754-761.
Zechmeister, E. B. (1972). Orthographic distinctiveness as a variable in word recognition. American Journal of Psychology, 85, 425-430.
Zechmeister, E. B., Curt, C., & Sebastian, J. A. (1978). Errors in a recognition memory task are a U-shaped function of word frequency. Bulletin of the Psy- chonomic Society, 11, 371-373.
[1] RAAM stands for Recursive Auto-Associative Memory a device introduced by Pollack (1988).
[2] I would like to thank an anonymous reviewer for suggesting this possibility.
[5] The Minnesota Daily articles are available for ftp on the internet at http://www.daily.umn.edu/.
Submitted 11th August 1995
Published 15th November 1995