by James Reason
6.1.2. Similarity-matching
According to William James (1908, pp. 118-119): “Whatever appears in the mind must be introduced.... This is as true of what you are recollecting as it is of everything else you think of.” This section is concerned with the way in which these ‘introductions’ are effected in question-answering: by the matching of the calling conditions present in a question to the attributes of knowledge items stored in semantic memory.
Similarity-matching is a recurrent theme across a wide variety of memory theories. It is the primary basis of memory search. The view adopted here is that the form of a general knowledge question delivers a set of retrieval cues to long-term memory (the knowledge base). These cues then automatically activate stored items possessing attributes that match either wholly or partially the ‘calling conditions’ communicated by the question. This is assumed to be an extremely rapid and efficient process, once the cues have been set loose within the knowledge base.
How these calling conditions are perceived plays a critical part in determining the course of the subsequent act of retrieval. For example, many Britons, when encouraged to persevere with the question ‘Who was the tall watcher on the tor?’, produce one of two sensible answers, depending upon how they interpret the available cues. One group correctly responds with Sherlock Holmes, the other with Heathcliffe. Both groups assume from the linguistic style of the question that it refers to a dated English literary figure; the difference in outcome hinges on how they understand the term ‘tor’. Those that say Sherlock Holmes recognise that ‘tor’ usually refers to a rocky outcrop on Dartmoor, from which they make the connection to Sherlock Holmes in the Hound of the Baskervilles. The others interpret ‘tor’ as a hilly place on any moor and conjure up the ready image (among educated Britons at least) of the brooding Yorkshireman (or of Olivier, his screen counterpart) in Wuthering Heights.
These reactions to the ‘watcher’ question also illustrate a further point: not all of the available calling conditions need be active at the outset. The more subtle clues to an answer may only be uncovered as the result of sustained inferential activity. Thus, a critical calling condition could remain concealed for some time, and only when its possible relevance has been established will it be loosed upon its automatic matching quest. Many questions will release all of their calling conditions immediately, but others will leak them only gradually into the retrieval system.
6.1.3. Frequency-gambling
The calling conditions supplied by a general knowledge question frequently fail to find an immediate right answer, either because the retrieval cues are insufficiently specified or because the relevant stored knowledge is incomplete. (These two possibilities are treated as being functionally equivalent within the model.) As a result, a number of partially matched ‘candidates’ can be activated at various stages of the search. The essence of frequency-gambling is that selection from among these candidates is biased in favour of the more frequently-encountered items, and that these are more likely to be called to mind for additional serial processing. Selection is necessary because the workspace available to handle the products of parallel search is extremely limited in its capacity.
The extent to which frequency-gambling plays a part in the retrieval process depends upon the quality of the match between the calling conditions and a specific knowledge item or restricted set of items. The less precise the ‘description’ of a stored fact provided by the calling conditions at the outset of the search and the more patchy the relevant knowledge held in memory, the greater will be the involvement of frequency-gambling. This is simply a reiteration of the underspecification principle stated earlier.
6.1.4. Inference
As used here, the term ‘inference’ embraces a number of different serial search procedures. Collectively, they involve the directed application of selected pieces of knowledge and various operators to the products of parallel search. Similarity-matching and frequency-gambling are the automatic processes by which various candidates are brought to mind. They operate in the realms of unawareness beyond the “intensively active gap” in consciousness described by William James (1890). Inference, on the other hand, refers to the work done upon these items within the limited workspace, once they have been ‘called to mind’.
In regard to the nature of this ‘mindwork’, it has been suggested that inferences fulfill two general functions (Warren, Nicholas & Trabasso, 1979). First, they establish connections between the propositions available in a question (calling conditions) and stored knowledge items. This not only guides the retrieval process, it also allows for the integration of both new and old information into a coherent body of knowledge. Second, since knowledge is always incomplete, inferences help the memory user to fill in the missing pieces (see Bartlett, 1932; Schank, 1982).
Even when people feel certain that a particular item of knowledge is not available in memory, they can sometimes piece together what they do not know by making plausible inferences from other kinds of stored knowledge. Consider the following question: “Which president of the United States was the first to travel in an aeroplane?” Although this is hardly a well known fact, most of us can make an educated guess at who it was, even though we may possess very little relevant knowledge. Our inferences might go along the following lines: The first manned flight was in the early 1900s; but until the end of the First World War aircraft were far too small and precarious for a president to be allowed to fly in one. Commercial aviation began in earnest sometime in the early 1920s. Who were the presidents of that time? Wilson, Harding and Coolidge. Wilson was probably too sick to fly even if there were aircraft large and safe enough to carry him and his retinue. So the most likely candidates are either Harding or Coolidge. Since presidents, being politicians, are quick to seize upon new ways of reaching the voters, it was probably Harding rather than Coolidge.
This may not be the right answer (which is still unknown to the writer), but the way it was reached shows how a feasible guess can be constructed by doing ‘inference work’ on scraps of knowledge drawn from two fairly disparate domains: aviation history and U.S. presidents. Although this chain of inference was presented as if it were a wholly conscious process, it will actually have involved several intermittent bursts of nonreportable parallel search: for the approximate date of the first manned flight, for the condition of early aircraft, for the beginnings of commercial aviation, for the presidents of the early to mid-1920s and so on. Thus, what might be recounted by the searcher as a coherent linear sequence is, in reality, a complex cycling of (a) items thrown up by unconscious retrieval mechanisms and (b) conscious inferential work, with the latter yielding fresh calling conditions for the former. Only the products of the parallel activity and a certain amount of the serial problem-solving are accessible to introspection.
6.2. Serial and parallel search processes
The distinction made above between the parallel and serial aspects of memory search corresponds in many important respects to that made by other theorists. For example, it has been argued (Atkinson & Juola, 1973; Mandler, 1980) that memory representations may be accessed in two ways: the one direct and rapid; the other slower and indirect, involving a laborious search along relatively unfamiliar associative pathways. Since these two search mechanisms differ in their demands upon conscious attention, Klatzky (1984, p. 75) suggested that they can be described in terms of their positions along an attentional continuum: “as a search relies more on nonhabitual, novel associations in the memory structure, it demands more attention. Direct search is achieved by activation of well established memory pathways and lies at the low end of the continuum. Associative search uses attention in varying amounts.” This proposal avoids the need to make hard and fast distinctions between conscious and unconscious processes, which are in any case hard to substantiate.
6.3. “The buck stops here”
Who said (or, more accurately, had a sign on his desk saying): “The buck stops here”? The answer is Harry S Truman. If this quotation is not in
your Truman ‘file’ (i.e., a stored attribute of Truman), then you might reason along the following lines. Quotation questions usually relate to famous people. The word ‘buck’ suggests an American. The most famous Americans are presidents. In three short and largely intuitive steps, the possible search set has been reduced from many millions to the (then) 39 American presidents. But which president?
The retrieval theory outlined above would predict that people who did not actually know the answer, but were willing to guess, would attempt to make a match with the first president that came to mind, probably the then incumbent, Ronald Reagan. If this did not ‘feel right’, they might then proceed down the frequency-of-encounter gradient... Kennedy, Carter, Nixon, Franklin D. Roosevelt, Lincoln, Washington, Ford ... until they came to a name that seemed to match the quotation.
If one asked this question of a large number of people who knew little of American presidents, it should be possible to obtain some idea of the relative influences of the various search strategies in providing the guessed answer. If only frequency-gambling prevailed, then their attributions would correspond precisely to the frequency gradient. If other processes also played a part, then the attributions would deviate in some degree from this correspondence, with less salient presidents receiving a larger number of attributions than the more frequently encountered ones. This was what actually happened when I put the question to 126 British psychology students who, on average, knew the names of only 9 of the 39 presidents. The results are shown in Figure 4.4.
Figure 4.4. Two distributions are plotted: p(re-call), the probability of recall of 23 American presidents; and p(attribution), the probability of attributing the “buck” quotation to a particular president. In descending p(recall) order, the first eight presidents are: Reagan, Kennedy, Carter, Nixon, FDR, Lincoln, Washington and Ford. Truman (the correct answer) is number l2.
In this case, the ‘salience gradient’ (the rear distribution) was derived from the probability of a president being recalled when the students were asked to generate as many exemplars as possible of the category ‘American president’ in five minutes (see also Roediger & Crowder, 1976). The quotation attributions are shown by the front distribution. The first point to notice is that over 80 per cent of these attributions were assigned to the five most salient presidents. The second point is that the number of attributions per president does not correspond to the salience order. Nixon received more than his fair share. Further probing yielded the following explanation. The subjects did not equate ‘buck’ with responsibility (not knowing the story of Mississippi gamblers and buck-handled knives); instead, they took it to mean a dollar. While they knew relatively little of the facts of Richard Nixon’s presidency, many of them thought he had been involved in some kind of shady dealing. On this basis, they interpreted the quotation to mean ‘The buck (the dollar) stops here (in my pocket)’, and Nixon as the likely source.
Although inaccurate, these assumptions clearly indicate the involvement, in certain subjects at least, of a matching process that overrode the dictates of the frequency gradient. Taken as a whole, this simple demonstration reveals something of the complex interaction between the three search mechanisms: similarity-matching, frequency-gambling and inference. We will explore further aspects of this interaction below.
6.4. The quotations study
This section describes a recent Manchester study (Reason, Horrocks & Bailey, 1986) designed to test the following prediction from the cognitive underspecification hypothesis: The amount of frequency-gambling (i.e., the tendency to emit high-frequency candidates) evident in the responses to general knowledge questions will be inversely related to the degree of relevant knowledge possessed by the respondent. The knowledge domain was American presidents, and the study was carried out in two separate stages: a mapping phase and a test phase.
The mapping phase was in two parts. In the first, two separate groups of psychology undergraduates (totalling 126 subjects) were asked, on separate occasions, to recall as many presidents as they could in a 5-minute period. In the second part, again on separate occasions, they were given a list of all (the then) 39 presidents and were required to make frequency-of-encounter ratings (on a scale of one to seven) for each one. These lists were in random rather than chronological order.
In the test phase, a questionnaire was given to 114 undergraduates, none of whom had been involved in the mapping phase. Each questionnaire was in two main parts:
(a) A recognition test, comprising the names of the (then) 39 incumbents mixed together with those of 78 famous contemporaries. These were in random order and the subjects were asked to underline the names of those presidents they could identify as such. The score, corrected for guessing, was assumed to give an approximate indication of the individual’s domain knowledge.
(b) Multiple-choice quotation attribution, in which four quotations were presented, each with an attribution selection of eight presidential names beneath (one of whom was the true source). For each quotation, subjects were required simply to underline their ‘best guess’ as to the source. The dependent variable was the frequency score for the selected president in each case. These scores were derived from the ratings obtained in the mapping phase.
The subjects were split into approximately equal quartile groupings according to their recognition scores—indicating their degree of domain knowledge. As predicted, the results revealed a highly significant inverse relationship between domain expertise and the tendency to frequency-gamble. The more people knew, the less inclined they were to attribute quotations to contextually appropriate, high-frequency presidents.
A more immediate ‘feel’ for these data is provided by Figure 4.5, which shows the pattern of attributions for the ‘low knowledge’ (LK) and ‘higher knowledge’ (HK) groups (a median split) in relation to the Teddy Roosevelt quotation: “A man who is good enough to shed his blood for his country is good enough to be given a square deal afterwards. More than that no man is entitled to, less than that no man shall have.”
The implicit cueing directs subjects to consider end-of-war or postwar presidents. In both groups, the majority guessed the correct source, but probably for the wrong reasons: a confusion between Franklin and Theodore Roosevelt that was further compounded by the similarity between Teddy’s ‘square deal’ and Franklin’s ‘New Deal’. Other data indicate that FDR is known by these undergraduates to be a wartime president, so, even allowing for the confusion, it is an intelligent choice. Of more interest, however, was the difference between the groups in the Kennedy attributions. To Manchester psychology students, John F. Kennedy remains a very salient president, being second only in rated frequency-of-encounter to Ronald Reagan. But since his term of office was some 15 years after World War Two and nearly a decade after the Korean conflict, he is not a particularly good match to the calling conditions supplied by the quote. For the LK group, therefore, he was an ‘ignorant’ guess; whereas the popular choices among the HK group reveal a fairly ‘educated’ pattern of guessing. Roosevelt (whichever one), Truman, Grant and Nixon all had good cause to be concerned about the welfare of veterans.
Figure 4.5. Comparing selections for the low- and high-knowledge subjects in regard to the “shed his blood ...” quotation. Most interesting are the differences for JFK between the two groups (see text).
6.5. The presidential recognition study
The underspecification hypothesis makes a number of predictions regarding the relative usage of similarity-matching and frequency-gambling. It postulates that similarity-matching will be the predominant strategy when (a) the ‘calling conditions’ (retrieval cues) are adequate to specify a unique knowledge item and (b) when there is a large number of stored items. In contrast, frequency-gambling will predominate (a) when the cues are insufficient or ambiguous and (b) when there is impoverished domain knowledge (i.e., low expertise). Thus, domain knowledge and cue sufficiency are seen as functionally equivalent in their effects upon the degree of retrieval underspecificati
on; a degradation of one or both would decrease the specification of the knowledge item, and would therefore increase the level of frequency-gambling.
In the quotation study, we were able to test the knowledge strand of this hypothesis. I found, as predicted, that the less subjects knew about American presidents, the more likely they were to attribute the quotes to high-frequency presidents. However, the use of quotations as test material did not permit the systematic manipulation of the specificity of the retrieval cues. A second study was therefore conducted that allowed us to control both domain knowledge and cue specificity.
As before, domain knowledge was assessed by the recognition test (described earlier). In this case, however, we sought to sample a much wider range of expertise than before, and so both American and British students were tested. There were 112 subjects, divided into three groups: low knowledge (who recognised fewer than 10 presidents); medium knowledge (who recognised between 10 and 30 presidents) and high knowledge (who recognised more than 30 presidents).
The test material was a two-part questionnaire, where both parts were constructed on the same basic principles. At the top of each page were listed a variable number of retrieval cues, each one relating to a biographical fact covering a variable number of presidents. Below were listed the same 20 presidents, but in a different order for each retrieval task. These were the 20 most frequently-encountered presidents (as shown by the ratings obtained from the ‘mapping’ phase of the quotation study), excluding Ronald Reagan and George Washington. In Section A, these biographical facts were unique to a single president. In Section B, they were possessed by a number of presidents (more than one, but fewer than 11). In both cases, the subjects were simply asked to underline those listed president(s) who best fitted the retrieval cues. In addition, they were required to rate their confidence in the selection (0-100%), and to indicate (in percentage values) what strategies they used in making their choice. They were offered four possibilities: pure guesswork, process of elimination, direct identification and other means (none used this last category).