The present study investigated, within the framework of spreading activation model, whether or not nodes are activated in a hierarchical fashion in the memory network. If, for example, activation of plant spreads to flower and then to rose, subjects should take longer to detect the relation between plant and rose than that between either plant and flower or flower and rose. The results of an experiment conducted to test the above hypothesis showed a significant difference between subjects' RTs for the three kinds of word pairs. Appropriate post-hoc comparisons of the mean RTs supported the hypothesis of this study. That is, the time taken for subjects to detect the relation between pairs such as plant-rose was significantly greater than the time taken for them to detect the relation between the words in either plant-flower or flower-rose pair types.

    Several studies have shown that spreading activation may be suggested as the underlying mechanism involved in such tasks as category exemplar production  and sentence verification (Loftus, 1974), episodic sentence and word recognition (Anderson, 1983a, 1983b), and perceptual word recognition (McCleland & Rumelhart, 1981).

    This spreading activation model also best accounts for an unconscious process called associative priming (Ratcliff & McKoon, 1981). Associative priming refers to the facilitation in access to information when associated items are presented (Anderson, 1995). For example, words are named faster in the context of an associated word (e.g., bread-butter) than in the context of an unassociated word (e.g., table-window) and also a word is recognized more quickly when it is preceded by a word from the same sentence or proposition (e.g., run-fast) than when it is preceded by a word from a different sentence or  proposition (e.g., common-fast).

    Spreading activation model accounts for this phenomenon by assuming that words which are semantically associated with each other are represented in the form of a network and activation spreads through this network from presented words to their associated words in memory (Anderson, 1995).

    Empirical evidence has shown that spread of activation is automatic as opposed to being under strategic control (Balota, 1983; Neely, 1977). It has also been shown that the amount of activation of a concept node is a function of the length of the associative pathway between the node and the source of activation (Lorch, 1982). Moreover, the amount of activation spreading from a given node along a pathway is a function of the strength of that pathway relative to the sum of the strengths of all paths emanating from that node (Reder & Anderson, 1980).

    One of the assumptions of  the spreading activation model is that, since concepts are assumed to be associated within a network of associations, activation may spread not only to directly related concepts but also from those concepts to concepts further in the memory network. This assumption has been known as "multiple-step" assumption. This assumption has been used as an explanatory construct for a number of memory retrieval phenomena, such as category verification (Collins & Quillian, 1969), and episodic sentence verification (Anderson, 1976).

    However, based on a series of experiments, de Groot (1983) suggested that activation spreads only a single step within the memory network (i.e., "one-step" activation). She constructed a set of triads in which  there was a direct relation between the first and the second word (e.g., bull-cow) and the second and the third word (e.g., cow-milk) but no direct relation between the first and third word (bull-milk). de Groot argued that if subjects were able to make a lexical decision to milk more quickly in the mediated prime condition (bull-milk) than in the neutral prime condition (blank-milk), then this would suggest that activation had spread across two associative pathways from bull to cow to milk and this would provide evidence for multiple-step spreading activation. Based on her findings, de Groot suggested that activation spreads to directly related concepts but does not spread any further within the memory network.

    Balota and Lorch (1986), however, doubt the validity of the conclusions reached by de Groot. One of the points they make is that the theories of spreading activation predict less priming facilitation in a mediated priming condition (bull-milk) than in a related priming condition (cow-milk). This prediction is based on the assumption that the amount of activation available at a node depends on its distance from the source of activation. Given the relatively small priming effects de Groot reports for directly related concepts (26 ms in Experiment 1), the amount of facilitation expected for the mediated condition would be quite small. Thus, with de Groot's materials, considerable power would be required to detect the small predicted effect in the mediated condition (see Balota & Lorch 1986 for the full criticism).

    Balota and Lorch (1986) also investigated whether activation automatically spreads beyond directly associated concepts within the memory network. In a series of lexical decision and pronunciation experiments they constructed prime-target pairs such that there was a relation between the prime (e.g., lion) and the target (e.g., stripes) only through a mediating concept (e.g., tiger). The lexical decision results they obtained yielded facilitation of directly related priming conditions (e.g., lion-tiger and tiger-stripes). However, the mediated condition (e.g., lion-stripes) did not facilitate performance compared to either a neutral prime or an unrelated prime condition. In contrast, the pronunciation results yielded facilitation of both directly related and mediated priming conditions. Thus, their results support the notion that activation spreads beyond directly related concepts in semantic memory.  As for the results of the lexical decision task, they suggested that the task had masked the appearance of a mediated priming effect.

    A point which needs to be made here about the results of the lexical decision tasks in de Groot as well as Balota and Lorch studies is that the kind of hierarchical relationship between the nodes represented by the words used in these studies is semantic. However, lexical decision task is not a powerful tool to tap semantic level. Thus, the experiments  which have failed to support the multiple-step hypothesis may have suffered from a methodological flaw. To eliminate this possibility, a new task was employed in the present study. The task was detection of relation, in which the subjects had to indicate whether or not the pairs are semantically related.

    In the present study we tested, with a novel design, the assumption that concepts are represented in a hierarchical relationship with each other in the memory network and in detecting the relation between pairs such as plant and rose, a mediating node (i.e., flower) is also activated. If this turns out to be the case, the prediction is that it should take longer for subjects to detect the relation between  plant and rose than either the relation between plant and flower or flower and  rose.

Apparatus. Both stimulus presentation and data collection were controlled by a 33 MHz PC, with a full color VGA monitor. Subjects made related/unrelated responses by pressing either the "1" or "2" key on the keyboard.

Materials. The critical stimuli were based on a set of 6 word triads. In each triad, the first and second words were directly related (e.g., plant-flower). The second and third words (e.g., flower-rose) were also directly related . The first and third words (i.e., plant-rose) were, however,  supposedly related through the second word (i.e., flower). That is, the second word (i.e., flower) formed a mediating concept between the first and the third one. The above-mentioned pairs were labeled (1-2), (2-3), and (1-3) as Pair-type factor (P-type) consecutively. On the whole, 18 pairs of words were constructed from the triads, six pairs from each type of relation.

     In addition to the critical pairs, 28 false word pairs, including 10 pairs of related words (e.g., cat-dog) and 18 pairs of unrelated words (e.g., doll-car) were constructed. The difference between these related words and the ones made using the triads was that the related ones (e.g., cat-dog) did not form a hierarchical relationship with each other, whereas the ones in the triads formed a hierarchical relationship.

    After all the stimuli were constructed, they were assigned to a list. Thus, the experimental list included 46 pairs of words. All the words in the list were from Persian, the native language of the subjects. The stimuli were all similar in length and frequency. An attempt was also made not to select  stimulus words which are semantically ambiguous. As the stimuli were all from Persian, the possibility of syntactic ambiguity was automatically ruled out. This is due to the fact that in Persian different parts of speech are made distinct from each other by the affixes attached to the words.

    The stimuli were all presented on the computer screen in white against a black background. The order of presentation of the pairs in the list was randomized. No experimental word pair was included among the first three pairs of the list.

Procedure.Subjects were instructed that they would be presented with a pair of stimuli on each of the 46 trials and they had to indicate whether or not the words in each pair were semantically related. They had to indicate their responses by pressing the "1" key if the words in the pair were related and the "2" key if the words were unrelated.

     The exact sequence of events on each trial was as follows: (a) a brief tone was presented for 1 s; (b) a fixation point (x) was presented for 3 s in the center of the screen; (c) the first word was presented for 300 ms (for the 320 ms SOA ); (d) a dark interval was presented for 20 ms; (e) the second word was presented for 300 ms; (f) a blank screen was presented until the subject pressed the "1" or "2" key; (g) the message, "Press any key to continue",  was presented until the subject pressed any key to start the next trial.

    Reaction times (RT) were measured from the disappearance of the second stimulus word to the time when a subject pressed the response button. Subjects participated individually and throughout an experimental session the subjects were seated comfortably approximately 40 cm from the monitor.

Data Analysis and Results
 Mean RTs and percentages of errors for the experiment are presented in Table 1. Errors and RTs in the relation detection task greater than 2000 ms and less than 200 ms were excluded from the RT analysis. The RTs for the false word pairs (n = 28) were also removed.

    The data were next submitted to a repeated-measure ANOVA on P-type to compare the subjects' mean RTs for each pair type (i.e., 1-2, 1-3, 2-3). The results show a significant difference in the means of the group RTs, F(2,58) = 4.42, p < .05. To make further comparisons among mean RTs, protected t procedures were employed, the results of which are presented in Table 2.

TABLE 1.
Mean latencies as a function of pair type. (Error rates are in parentheses)
_________________________________________________________
                            Pair Type
    _________________________________
                        1-2        1-3             2-3
_________________________________________________________
     RT            463        553.2         482
                    (2.5%)    (4.0%)      (.9%)
__________________________________________________________
 
 

TABLE 2.
Post-hoc comparisons for the mean Rts of different pair types.
___________________________________
 Pair Types Compared          t value
________________             ______
        (1-3)/(1-2)                     2.81**
        (1-3)/(2-3)                     2.22*
        (2-3)/(1-2)                     .59
___________________________________
**p < .01, *p < .05, df = 58
 
 

 Network models of the human memory maintain that nodes in the memory network are represented in a hierarchical fashion, and that activation also spreads in a hierarchical fashion in the network, that is, from a node to an immediately lower node and then to a node still lower in the hierarchy. For instance, activation of plant would first spread to flower and then to rose.

     One of the major assumptions in spreading activation model is that the amount of activation reaching a node depends on the distance from the source of activation. Thus, the detection of relation in the case of plant-rose should take longer than those of either plant-flower or flower-rose. The results of the present experiment support the above-mentioned assumptions.

    In memory network, activation reduces as it traverses an intermediate node, because the amount of activation emanating from any node in the network is proportional to the strength of all pathways emanating from that node. Thus, we can not expect the amount of activation spreading from node 1 to node 3  to be twice as much the amount of activation spreading from either 1 to 2 or from 2 to 3. The results of the present study also seem to support this assumption (see Table 1).

Anderson, J. R. (1976). Language, memory and thought. Hillsdale, NJ: Erlbaum.

Anderson, J. R. (1983a). The architecture of cognition. Cambridge, MA: Harvard University Press.

Anderson, J. R. (1983b). A spreding activation theory of memory, Journal of Verbal Learning and Verbal Behavior, 22,
    261-295.

Anderson, J. R. (1995). Cognitive psychology and its implications (4th ed.). New York: W. H. Freeman.

Balota, D. A. (1983). Automatic semantic activation and episodic memory  encoding.Journal of Verbal Learning and
    Verbal Behavior, 22, 88-104.

Balota, D. A., & Lorch, R. F.(1986). Depth of automatic spreading activation:  Mediated priming effects in pronunciation but
    not in lexical decision. Journal of Experimental Psychology: Learning, Memory, Cognition, 12,  336-345.

Collins, A., and Quillian, M. (1969). Retrieval time from semantic memory.  Journal of Verbal Learning and Verbal
    Behavior, 8, 240-248.

de Groot, A. M. B. (1983). The range of automatic spreading activation in word priming. Journal of Verbal Learning and
    Verbal Behavior, 22, 417- 436.

Loftus, E. F. (1974). Activation of semantic memory. American Journal of  Psychology, 86, 331-337.

Lorch, R. F. (1982). Priming and search processes in semantic memory: A test  of three models of spreading activation.
    Journal of Verbal Learning and Verbal Behavior, 21, 468-492.

McClelland. J. I., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. an
    account of basic findings. Psychological Review, 88, 375-407.

McNamara, T. P., & Diwadkar V. A. (1996). The context of memory retrieval. Journal of Memory and Language, 35,
    877-892.

Neely, J. H. (1977). Semantic priming and retrieval from lexical memory: Roles of  inhibitionless spreading activation and
    limited capacity attention.Journal of Experimental Psychology: General, 106, 226-254.

Ratcliff, R., & McKoon, G. (1981). Does activation really spread? Psychological Review, 88, 454-462.

Reder, L. M. & Anderson, J. R. (1980). A partial solution of the paradox of  interference: The role of integrating knowledge.
    Cognitive Psychology, 12, 447-472.

Roelofs, A. (1992). A spreading-activation theory of lemma retrieval in speaking. Cognition, 42, 107-142.