The influence of item properties on association-memory

https://doi.org/10.1016/j.jml.2010.03.001Get rights and content

Abstract

Word properties like imageability and word frequency improve cued recall of verbal paired-associates. We asked whether these enhancements follow simply from prior effects on item-memory, or also strengthen associations between items. Participants studied word pairs varying in imageability or frequency: pairs were “pure” (high–high, low–low) or “mixed” (high–low, low–high) where “high” and “low” refer to imageability or frequency values and are probed with forward (A–?) and backward (?–B) cues. Probabilistic model fits to the data suggested that imageability primarily improved retrieval of associations, but frequency primarily improved recall of target items. All pair types exhibited a high correlation between forward and backward probe accuracy, a measure of holistic learning (Kahana, 2002), which extends the boundary conditions of holistic association-memory and challenges Paivio’s (1971) suggestion that holistic learning depends critically on imagery. In sum, item properties can boost association-memory beyond simply boosting target retrievability.

Introduction

Much of everyday memory function involves deliberate retrieval of information associated with cues in the environment. For example, meeting a friend cues your memory (cued recall probe) causing you to recall her husband’s name (cued recall target). This kind of memory requires both retrieval of an association (relationship) and production of a target item (item recall). This kind of memory function is typically studied with cued recall of paired-associates (Calkins, 1896, Underwood, 1966). The participant is required to learn and retrieve the association. However, cued recall performance could be influenced by memory for the association between the items in the pair, as well as by memory for the individual items themselves (e.g., Hockley & Cristi, 1996). Some material types are remembered better than others in cued recall, and although one would be tempted to attribute such differences to effects of the underlying encoding and retrieval of associations, it is also possible that such effects can be entirely attributed to differences in retrievability of the target items. Consider two properties of words that are known to improve cued recall accuracy: frequency of usage, or word frequency (Clark, 1992, Clark and Burchett, 1994) and subjective ratings of conduciveness to mental imagery production, or imageability (Lockhart, 1969, Paivio, 1965, Paivio, 1968, Paivio et al., 1968, Wood, 1967). Both of these item properties have well established effects on memory for items: high-frequency words are better recalled but worse recognized (the “word frequency paradox”; e.g., Gorman, 1961, Gregg, 1976, Hall, 1954, Hall, 1979, Shepard, 1967) and high-imageable words are better recalled and better recognized (e.g., Gorman, 1961). It is possible that the cued-recall advantage for high-frequency words and high-imageable words follows directly from the greater retrievability of target items. Alternatively, high-frequency words and high-imageable words may in fact lead to better encoding and retrieval of the associations between paired items.

Our first objective was to determine the locus of item-property effects in cued recall performance. Specifically, item-property effects could be related to memory for associations or memory for the items themselves. Our second objective was to ask whether these item properties influence the relationship between forward and backward associations, a measure that is diagnostic of holistic association-memory (Kahana, 2002). Our third objective was to directly test Paivio’s (1971) conceptual-peg hypothesis, which suggests that pairs consisting of two low-imageability words cannot be learned holistically as the association cannot be ‘pegged’ to an image suggested by either word.

A related question has been asked about the effect of word frequency on serial recall and free recall with respect to whether word frequency improves (a) memory for serial order or (b) the simple retrievability of the list items (Hulme et al., 2003, Ward et al., 2003). These studies compared memory for pure lists with memory for lists that alternated between items of high and low-frequency. The obvious sign of enhancement of item recall alone would be a zig-zag pattern wherein alternating lists literally alternate between the pure-high and pure-low accuracy levels. However, in serial list learning, manipulations of word frequency have suggested that the enhanced memory for high-frequency words results from improved serial-order memory (e.g., item–item association strengths) rather than from item recall.

We follow a similar logic to disentangle item- versus association-memory effects in cued recall of pairs. In Experiment 1a, participants studied pairs composed of words that had high versus low-imageability values or high versus low-frequency values. For each manipulation (word frequency or imageability), pairs were either pure (composed of two items of the same class, i.e., high–high or low–low), or mixed (composed of items differing in class, i.e., high–low or low–high). Given a studied pair (AB) cued recall probes can be in the forward direction (A–?) or in the backward direction (?–B). Half the cued recall probes for each pair type were forward and half were backward. Subsets of our design have been carried out, but as some conditions were missing, these results are open to interpretation. Briefly, studies of imageability have found symmetric mean performance in both pure pairs and mixed pairs (Bower, 1972, Crowder, 1976, Paivio, 1971, Wollen and Lowry, 1971). Word frequency studies have produced symmetric mean performance in pure pairs, but asymmetric mean performance in mixed pairs (Crowder, 1976, Nelson and McEvoy, 2000, Paivio, 1971). The results of these studies suggest that imageability primarily acted through modulating the strength of the association, or through item retrievability. Word frequency was found to likely act through item retrievability. However, there is still some interpretational ambiguity as these studies not have sufficient conditions to make definitive conclusions or contrast all possible mechanisms. Our complete design allowed us to fit the data using a simple model to obtain separate estimates of the effects of each item property on association-memory and target-item recall probabilities.

The model was based on the assumption that accurate cued recall relies on two kinds of mechanisms: (a) successful recall is dependent on individual properties of the cued recall probe and target items, individually, and (b) successful recall is also dependent on retrieval of the relationship between paired items. Here we make the simplifying assumption that retrieval operations are independent; thus retrieval probabilities combine multiplicatively. The model estimates the degree to which properties of the probe, the target, and the relationship between items influence probability of recall. To understand how the model works, consider two extreme cases, depicted in Fig. 1. Fig. 1a presents simulated data generated by a model that treated high- and low-valued items and pairs identically except that it assumed that high-valued target words would be better recalled. In this example, the difference between pure high (HH) and pure low (LL) pairs is clarified by the asymmetries observed in the mixed pairs (HL and LH). Accuracy is high whenever the target item is high-valued, regardless of the identity of the paired (probe) item. In contrast, Fig. 1b was generated by a model that treated high- and low-valued items identically except that it assumed that as the number of high-valued words in a pair increased, probability of retrieving the association would increase, regardless of which item was the probe or target. In this example, the same level of cued-recall advantage for pure high versus pure low pairs is clarified by the mixed pairs, which show no asymmetries.

As an additional measure of item retrieval effects, we also decided to look at intrusion rates for the high and low items in cued recall. We hypothesized that if the high items are more retrievable than low items, participants will recall them more often as cued recall errors. However, if high and low items are equally retrievable, they should intrude in cued recall in equal proportions.

Separate from the question of whether associations are learned better or worse, one can ask whether material type can influence the holistic nature of the learned association. Consider the association between a pair of items, AB. One can decompose the association into two directional associations, a forward association, AB, and a backward association, AB. The Independent Associations Hypothesis is that the association between paired items, AB, is composed of two separate, unidirectional associations (Wolford, 1971). Accordingly, AB is learned in a statistically independent step from AB. The consequence is that performance for forward probes of a pair, A-? is expected to be independent of performance for backward probes of the same pair, ?–B. In contrast, the Associative Symmetry Hypothesis assumes that pairs are learned as a compound, holistic unit (Asch and Ebenholtz, 1962, Köhler, 1947). Thus, forward and backward probes should be sensitive to the same variability in learning.

Gestalt psychologists initially claimed that associative symmetry implies equivalent performance on forward and backward cued recall tests (e.g., Asch & Ebenholtz, 1962), a finding that has been replicated numerous times (see Kahana, 2002, for a review). However, Kahana (2002) pointed out that symmetry of mean performance is orthogonal to whether a pair is learned holistically. He argued that the direct support for holistic learning would be an observation of a high, near-unity correlation between forward and backward tests at the level of individual pairs, over successive tests. Holistic learning would be observed as a high correlation if forward and backward probes measure the same underlying associative strength. Indeed, such a high forward–backward correlation has been observed in several studies of verbal paired-associates learning (e.g., Caplan et al., 2006, Kahana, 2002, Rehani and Caplan, in preparation, Rizzuto and Kahana, 2000, Rizzuto and Kahana, 2001) and object–location learning (Sommer, Rose, & Büchel, 2007).

To understand why the mean performance and correlation measures reflect different memory phenomena, consider a hypothetical cued recall experiment using a brief study list consisting of two pairs, SHROUD–RUMOUR and HELMET–MALICE. A participant might recall SHROUD–? and ?–MALICE correctly, but ?–RUMOUR and HELMET–? incorrectly. Mean performance for forward and backward probes is equal (both 50%) but the tests are not positively correlated, which is suggestive of non-holistic learning (Independent Associations Hypothesis). However, another participant recalls SHROUD–? and ?–RUMOUR correctly but HELMET–? and ?–MALICE incorrectly. Here, mean performance is still symmetric (50% for both forward and backward probes); however forward and backward probe performance is correlated at the level of pairs, suggesting holistic learning (Associative Symmetry Hypothesis). In other words, forward and backward probes test the same learned information.

What is interesting is that while this illustrates that the two measures (correlation versus mean performance symmetry) are mathematically independent, it is not known whether they are separable in human behaviour in tasks of associative learning (e.g., asymmetric mean performance with no reduction in forward–backward correlation). In previous research, Caplan, 2005, Caplan et al., 2006 partly dissociated these two measures in serial lists. Rehani and Caplan (in preparation) found symmetry in mean performance of pairs with reduced correlations. In the present study, we included successive testing (each pair tested twice to examine the relationship between forward and backward cued-recall performance) and asked whether the two measures are coupled empirically. We aimed to induce asymmetries in mean performance via mixed pairs, as has been achieved by prior studies (e.g., Ebbinghaus, 1885/1913, Horowitz et al., 1966, Lockhart, 1969, Paivio, 1965, Paivio, 1968, Paivio, 1971, Wollen and Lowry, 1971) and asked whether the forward–backward correlation is disrupted when mean performance is highly asymmetric.

One well developed theory of the influence of imageability on memory for verbal pairs is Paivio’s (1971) conceptual-peg hypothesis. A core assumption of this hypothesis was that imageability of words facilitates the participant’s ability to form an image combining the paired items, and that this image functions as a Gestalt. He further argued that if even one of the paired words is imageable (concrete), it can act as a ‘peg’ to which the low-imageable (abstract) item can be attached. However, if both words are abstract, no image can be formed. Because he assumed that the holistic representation was an image, the conceptual-peg hypothesis implies that pairs comprised of two low-imageable items, when learned, will not be learned holistically. Paivio and colleagues provided evidence based on measures of mean performance, which, as we elaborated in the previous section, cannot directly test whether or not associations are learned holistically (Kahana, 2002). We tested the conceptual-peg hypothesis by asking whether the correlation between forward and backward probes of pure abstract pairs was reduced compared to pairs containing high-imageable words (Experiments 1a and 1b).

To summarize the goals of the present study, we asked several questions regarding the effects of item properties on memory for associations. First, by presenting pure and mixed pairs of words simultaneously we asked whether single-item properties influence item-learning, association-learning, or both. We tested the hypothesis that these single-item properties influence not only recall of target items but also the recall of the associations between items. Second, when asymmetric mean performance is found, as expected in cued recall performance of the mixed pairs (Experiment 1a), this could signal a disruption of the holistic association. We tested whether greater asymmetry in cued recall implies a greater reduction in the correlation between forward and backward cued recall. Third, we tested the conceptual-peg assumption that the holistic representation is an image by asking whether pure low-imageable pairs exhibit reduced forward–backward correlation compared to high–high-imageable pairs (Experiments 1a, as well as a follow-up experiment, 1b).

Section snippets

Participants

Fifty-nine undergraduate students from the University of Alberta participated in the two-session study for partial fulfillment of an introductory psychology course requirement. Data from three participants were not included in our analyses because these participants failed to appear for the second session. Sixteen male and 40 female participants (mean age ± sd = 21.4 ± 4.7) were included in the analyses. Participants were required to have English as their first language, to have normal or

Cued recall accuracy

For both the word frequency and imageability sessions, repeated-measures ANOVAs were conducted on mean accuracy. Table 2 gives an overview of the pair types and their respective probe and target items.

Experiment 1b

Paivio (1971) held that low–low-imageability pairs would not be learned holistically. In Experiment 1a, we were unable to detect a disruption of holistic pair learning (using the successive testing measure), for the low–low-imageable condition. However, Sommer et al. (2007) found that for object–location pairs, if the presentation of the two items was sequential rather than simultaneous, which increasingly challenges the participant to link the paired items, it is possible to induce a

Results

The analysis methods for the second experiment were nearly identical to those used in Experiment 1a. However, because only pure pairs and one experimental session were used, the ANOVA had the design PAIRTYPE[2]×TESTDIRECTION[2]×TESTNUMBER[2]. Model fits could not be carried out as there were no mixed pairs.

Discussion

In two related experiments, we investigated the effects of two item properties, imageability and word frequency, on cued recall performance. Imageability primarily influenced memory for the associations, whereas word frequency primarily influenced target-item recall. Additionally, we replicated the finding of a high correlation between forward and backward cued recall accuracy, which follows from the notion that pairs are learned holistically, as opposed to forward and backward associations

Conclusion

Our findings elucidate the effects of item properties on cued recall in several ways: (a) Properties of single-items can either affect primarily item-memory (i.e., word frequency) or primarily memory for associations (i.e., imageability). (b) Single-item properties cannot by themselves disrupt the holistic-like association learning. (c) Models of association-memory must be able to accommodate overall highly correlated forward and backward associations and particular models may have multiple

Acknowledgments

We thank Chris Westbury for assistance in creating the word pools as well as Anthony R. McIntosh and Zainab Fatima for help with early planning of the experiments. Partly supported by Natural Sciences and Engineering Research Council (NSERC) of Canada.

References (72)

  • S.E. Asch et al.

    The principle of associative symmetry

    Proceedings of the American Philosophical Society

    (1962)
  • Baayen, R., Piepenbrock, R., & Gulikers, L. (1995). The CELEX lexical database (Release 2) [CD-ROM]. Philadelphia, PA:...
  • H. Bird et al.

    Age of acquisition and imageability ratings for a large set of words, including verbs and function words

    Behavior Research Methods, Instruments, & Computers

    (2001)
  • Y. Bishop et al.

    Discrete multivariate analysis: Theory and practice

    (1975)
  • G.H. Bower

    Imagery as a relational organizer in associative learning

    Journal of Verbal Learning and Verbal Behavior

    (1970)
  • G.H. Bower

    Mental imagery and associative learning

  • K.E. Burnham et al.

    Model selection and multimodel interference

    (2002)
  • K.E. Burnham et al.

    Multimodel inference: Understanding AIC and BIC in model selection

    Sociological Methods & Research

    (2004)
  • M.W. Calkins

    Association: An essay analytic and experimental

    The Psychological Review

    (1896)
  • J.B. Caplan et al.

    Linking associative and list memory: Pairs versus triples

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2006)
  • K.A. Chalmers et al.

    Experimental manipulation of prior experience: Effects on item and associative recognition

    Memory & Cognition

    (2003)
  • S.E. Clark

    Word frequency effects in associative and item recognition

    Memory & Cognition

    (1992)
  • S.E. Clark et al.

    Word frequency and list composition effects in associative recognition and recall

    Memory & Cognition

    (1994)
  • S.E. Clark et al.

    Global matching models of recognition memory: How the models match the data

    Psychonomic Bulletin & Review

    (1996)
  • S.E. Clark et al.

    Cuing effects and associative information in recognition memory

    Memory & Cognition

    (1992)
  • M.J. Cortese et al.

    Imageability ratings for 3,000 monosyllabic words

    Behavior Research Methods, Instruments, & Computers

    (2004)
  • R.G. Crowder

    Principles of learning and memory

    (1976)
  • R.A. Diana et al.

    The low-frequency encoding disadvantage: Word frequency affects processing demands

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2006)
  • Ebbinghaus, H. (1885/1913). Memory: A contribution to experimental psychology. New York: Teachers College, Columbia...
  • A.S. Geller et al.

    PyEPL: A cross-platform experiment-programming library

    Behavior Research Methods, Instruments, & Computers

    (2007)
  • A.N. Gorman

    Recognition memory for names as a function of abstractness and frequency

    Journal of Experimental Psychology

    (1961)
  • V.H. Gregg

    Word frequency, recognition and recall

  • J.F. Hall

    Availability and associative symmetry

    The American Journal of Psychology

    (1954)
  • J.F. Hall

    Recognition as a function of word frequency

    The American Journal of Psychology

    (1979)
  • C.A.G. Hayman et al.

    Contingent dissociation between recognition and fragment completion: The method of triangulation

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1989)
  • D.L. Hintzman

    Simpson’s paradox and the analysis of memory retrieval

    Psychological Review

    (1980)
  • Cited by (60)

    • Norming retrieval processes

      2020, Journal of Memory and Language
    • Order of items within associations

      2017, Journal of Memory and Language
    • Emotional arousal impairs association-memory: Roles of amygdala and hippocampus

      2017, NeuroImage
      Citation Excerpt :

      The results in the 5-AFC task closely resemble the previous cued recall results (Madan et al., 2012), namely, a reduction in association-memory for negative pure pairs compared to neutral pure pairs, with intermediate accuracy for mixed pairs but better performance for negative targets. Differences in associative memory accuracy (cued recall in Madan et al., 2012) for different materials can result not just from influences on the association-memory strength, but from effects on the item-level (see also Madan, 2014; Madan et al., 2010). As outlined in detail in Madan et al. (2012), our previous computational model formally tested whether association memory accuracy for negative compared to neutral information was influenced by item-level parameters (‘target retrievability,’ ‘cue effectiveness’) or by the association-memory strength itself.

    View all citing articles on Scopus
    View full text