part of the search process, the recall of the old names sets the context for the search for ..... One rummages about in the search domain picking out a piece that.
The Process of Retrieval from Very Long-Term Memory Xerox, PARC
University of California, San Diego
In this paper we argue that the protocols of subjects recalling the.names of their high school classmates, as well as an army of traditional memory phenomena, can be understood from an informationprocessing analysis which interprets ret r i e d as a problem-solvingprocess. This characterizationof retrieval focuses on the reconstructive and recursive nature of the process of remembering. Refrieval is viewed as a process in which some information about a target item is used to construct a description of the item and this description is used in attempts to recover new fragments of information. A new description is formed and the process continues recursively. Three subprocesses are identified: FIND A CONTEXT, in which a proper environment for conducting a search i s established: SEARCH, in whidr bits and pieces of information appropriate to the context are recovered until an adequate description can be formed within the search context; and VERIFY, in which recovery infonnation i s checked for consistency.
..
,
From the perspective d t h i s characterization, we examine the protocols of subpcts asked to think aloud while recalling the names of their classmates. Subiects' retrievals continued for a surprisingly long period and were characterized by the phenomena of overshoot, systematic hypothesizing, fabrications, establishment of search contexts, self-corrections, and the use of a number of basic search strategies. The view of memory as a problem-solving process not only leads us to the identification of these phenomena but also to a framework for o coherent interpretation for them.
INTRODUCTION
It is our view that remembering is a complex problem-solving activity precisely because humans are commonly constrained to ,paper.is
88
WILLIAMS AND HOLlAN
There are two primary threads to this presentation. One is a complex argument beginning from primary principles and traversing a succession of implications and transformations to yield the outline of what we believe to be a fundamentally different view of human memory. The second is an attempt to use a real-world recall task to serve as an exemplar of our evolving view of human remembering. This second theme is woven in our theoretical arguments and actually will precede, in part, their development. The theory is necessary to provide an interpretation of observations and the observations provide an instantiation of the theory. Our final product is not a new theory of memory. Rather, it is a characterization, an outline, of a theory. Our ultimate goal is a complete theory of remembering as a problem solving process, but we are not yet there. What we are able to provide is a primitive characterization of a reconstructive and recursive retrieval process and the outline for a mechanistic account of several novel and traditional memory phenomena. In outline our argument (diagramed in Figure 1) begins with a set of primitive principles: partial information (the central principle), descriptions, and large memory capacity. These principles suggest a kernel retrieval process which is the elemental process from which a global remembering process can be built. The use of this kernel process forces a set of problems upon any global retrieval process built from such elements. These problems are too little information, too much information, and false recoveries. These problems suggest a set of problem-solving tactics we expect to see in human recall. We find that these tactics can be observed almost directly in subjects' verbal protocols of thinking aloud while recalling the names of their classmates. We take these observations to provide support for our view, as well as to serve as examples to communicate ideas which otherwise remain perilously abstract. We then introduce a characterization of the retrieval process. This is an attempt to place the fundamental retrieval problems our view entails into an active form. We finally attempt to map our view of remembering onto some traditional memory phenomena such as forgetting, errors, judgments of certainty, and recall versus recognition. This is particularly difficult because our view holds that there are neither single nor simple causes for these phenomena. Rather, they find their origin and behavior coming from complex interactions within the dynamics of the process of remembering. Let us apologize in advance for the seeming redundancy of some of our discussion. Much of this redundancy is necessitated by the successive layers of our argument in which we carry the implications of partial information and other primitive notions through a multitude of theoretical and phenomenological transformations. To set the stage for our argument, let us begin by exploring recall in a natural situation in which subjects recall the names of their classmates from high school. Such recalls can be complex and convoluted tasks taking place over the
PRINCIPLES:
BOUNDARY OF ANALYSIS:
GLOBAL RETRIEVAL PROBLEMS:
RECURRENT PHENOMENA: (TACTICS)
PARTIAL DESCRIPTIONS LARGE MEMORY INFORMATION CAPACITY
KERNEL RETRIEVAL PROCESS
TOO MUCH TOO LITTLE INFORMATION INFORMATION
CONTEXTS
EXTENDED RETRIEVAL
SEARCH
SYSTEMATIC HYPOTHESIZING
FABRICATIONS
I
INFERENTIAL RECALL
SELFCORRECTIONS
FIND A CONTEXT
SEARCH
I RETRIEVAL CYCLE:
FALSE RECOVERY
Figure 1. Outline of argument.
course of weeks; Incidental details come to mind almost before the recovery of the first name. The high school buildings, the classes attended, the activities engaged-all seem t'o be a part of the process of retrieval. The following is a sample protocol from a subject asked to think aloud while attempting to recall the names oflis classmates:
..
'ihe first thing that comes to mind is . I m& it's almost l i e images of different snapshots of our high school. You know. I can think of our general science class, and waiting in the lunch line, and halls. Umm. Sort of, Jeff Thompson! [ ~ l names, l except those of public figures and a few cases where the phonetics of the name are at issue, have been changed in the protocols presented in this paper.] He was a friend of mine. Sort of pops into mind and I think, umm, we used to stand in lunch line together, and he was in our general science class. That's where I first met him, our freshman year. There was, umm. let's see I'm trying to think of people I interacted a lot with. And some them . . . are sort of people I've known after high school. Like
90
WILLIAMS AND HOLUN
.
Bill Newell. I . . I lived with him for a while, in Portland. Umm, after school. So he sort of comes to mind immediately too. Umm, let's see. I mean I guess it's almost easier for me to t h i i of our home town, and think of people . . . that I've still run into, on occasion, when I go back there. And then sort of check to see if they meet the requirements. Like were they in high school with me. And I can think of people l i e Buddy Collender, and John Tremble, who still both live in our . . . Ah . . . home I guess it's. It also seems that I want to think of, sort of, it's clear town. Umm, that I have to think of some other situations. It's like I want to think of, sort of prototypical situations and then sort of examine the people that were involved in those. And things like P.E. class, where there was . Ah. . . . Gary Booth. Umm, and Karl Brist, were sort of, we always ended up in the same P.E. classes, for some reasons. Umm, . I can think of things like dances. I guess then I usually think of girls (chuckle). L i e Cindy Shup, Judy Foss, and Sharon Ellis. I went to of grade school with her. Umm, . I mean it's sort of like I have a picture of the high There's a school dance. You know, and these are the women I knew then. Umm, girl I can think of, I have a very good picture of her but I can't remember her name. I can even remember she was, she was from Gales Creek. Ah . . And . Ah . . . I mean I can see what she looks like. I just can't remember her name. She lived near John Randolf, who was another friend of mine.
. ..
..
. ..
..
..
.
... . .
What is going on in a search of this nature? Do the pieces of incidental information about the school, or where someone lived, or the activities the subject engaged in serve some purpose? What guides this search? What is the relationship between searches of this kind and the vast array of laboratory phenomena which psychologists have studied for so long? These are some of the questions we are attempting to address. We have chosen this particular task for a number of reasons. First, it is something that almost everyone can do. Second, it has a definite standard of recovery, a collection of names. Third, we can calculate the validity of substantial portions of the recall by consulting the subjects' yearbooks. And finally, and perhaps most important, the task seems to make explicit a wide array of the phenomenology characteristics of difficult recall tasks and to provide a domain in which we can unfold the theoretical position we are developing.
METHODS
The basic approach taken in these studies is observational as opposed to experimental. Rather than manipulating some sets of variables to prove one hypothesis over another, we have chosen to collect a detailed set of observations. Subjects are asked to think aloud as they try to recall the names of their classmates. The data are the protocols that the subjects generate, including, of course, the names themselves and the times between names. In addition, at the conclusion of the experiment (usually after about 10 hours of recall spread over several weeks), we conducted extensive debriefing of the subjects, attempting to determine the origins of errors.
PROCESS OF RETRIEVAL
Procedure Four subjects were studied individually. They were instructed to think aloud while attempting to recall the first and last names of their classmates from high school. The yearbooks from each subject's high school were available to us and provided independent verification of the subjects' recollections. Subjects were run from 4 to 10 hours, in approximately one hour sessions. All sessions were tape recorded. Subjects were first asked a warmup question in order to insure that they understood the protocol instructions. The question was "Name the state capitals that begin with the letter 'B'." During their protocols, subjects were prompted whenever they had prolonged silences with comments from the observer such as: "What are you thinking now?" "What is passing through your mind?" etc. was one exception to the general procedure. Subject S4 was allowed to tape record her thinking aloud without the presence of the observer. Long silences did occasionally occur with ~ 4 . Subjects 1 S 1 and S3 were run in sessions of approximately one hour, one session per day, five days per week, for two weeks. Subjects S2 and S4 were run in sessions of from one to two hours, at more erratic intervals due to their personal, schedules. Subjects.weredirected to avoid thinking about their high school classmates between sessions, but this was, of course, not completely possible. Names were occasionally recalled inadvertently between sessions. Before each session,subjects were asked to report all the names they had recalled between sessions and to relate the circumstances of the recalls.
here
Subjects
The four subjects (S1 through S4) were from 4 to 19 years out of high school. All attended high school in the San Diego area. The subjects were paid $2 per hour for their participation in this experiment. Individually, their backgrounds are as follows:
Subject SZ; S1 is a female, five years out of high school. She has lived in San Diego in the same house for her entire life. She attended only one high school for two and a half years, graduating six months ahead of her class. Most of the subject's classmates from junior high also went to high school with her. Her graduating class had 609 members listed in the senior yearbook. Subject S2. S2 is a female, 19 years out of high school. She has lived in the San Diego area except for four years of college. At present, she has a home in the neighborhood of her high school. The subject attended only one high school and most of her classmates from junior high attended high school with her. Her graduating class had 3 18 members listed in the senior yearbook. S2 graduated in the same class as S4.
92
WILLIAMS AND HOLIAN
t
Subject S3. S3 is a female, four years out of high school. She has lived in several different regions of the San Diego area and spent her early years out of the state of California. She attended her f i t year in one high school, transferring to a second high school for her last two years. Most of her classmates from junior high attended a different high school than the subject. Her graduating class had 750 members listed in her senior yearbook.
-
Subject S4. S4 is a female, 19 years out of high school. She lived in the San Diego area until after graduation from high school. She left San Diego for around nine years, and returned about ten years ago. She was a member of S2's graduating class, is a personal friend of S2, and has maintained contact with S2 throughout the years. S4 attended only one high school and most of her junior high school classmates attended the same high school. Her graduating class had 318 members listed in the senior yearbook.
Scoring a Protocol The scoring of protocols is a complicated pmedure. It is not simply a matter of listing all the names recalled and verifying them against the subject's yearbook. Subjects expressed varying degrees of certainty about their recollections. They changed their minds (in some instances several times). On occasion, they had difficulty verifying the correcmess of a name. They mentioned names in passing which they did not intend to be candidates for names of high school classmates. Often subjects remembered nicknames of classmates which are difficult, and sometimes impossible, to verify. Sometimes subjects confused the name of one classmate with that of another. There are two basic categories of names mentioned: those mentioned as possible "candidates " for classmates and those mentioned "in passing. " An example of how a name might be mentioned in passing would be: Umm, well, my next impulse is to go to the art class where we all-our teacher's name was Bill Dane, uh. This was in a class where we related to people like human beings instead of other desks. where
...
In-passing names are operationally defined as those that the subject identifies as not being a candidate classmate either before or immediately after the name is mentioned. On occasion, identification as a noncandidate name is indicated by the context or tone in which the name is mentioned. All names which are not in-passing names were considered candidate names. Candidate names can be correct or incorrect as verified by the subject's senior yearbook. We refer to incorrect names as fabrications. Sometimes subjects recalled names that they had recalled before. Frequently, this is simply a part of the search process, the recall of the old names sets the context for the search for names not yet recovered. On occasion, however, subjects recalled a name unaware that the name had already been recalled. Indeed, the judgment of
93
PROCESS OF RETRIEVAL
whether a name has already been recalled is itself a recollection that can be correct or incorrect. Another scoring problem is the verification of nicknames. For scoring nicknames, we have taken the following policy: If there exists a verified classmate with a proper first or middle name which is commonly given the nickname the subject has specified (e.g., Bill for William, Patty for Patricia), then the nickname is scored as correct. If the subject later mentions the correct proper name in conjunction with the nickname (e.g., "Oh Bobo Richards is probably listed as Barbara Richards in our yearbook. That was her real name. "), then the nickname is scored as correct. If the nickname can be supported by other pictures or signatures in the yearbook (e.g., Kathleen Johnson is also listed as Nancy Johnson in a separate picture), then the nickname is also scored as correct. All other cases are scored as fabrications. The incorrect names, or fabrications, that subjects mention can be of a variety of types. The categories of fabrications are discussed in detail in a later section. Fabrications are difficult to score, in part, because subjects frequently introduced candidate names which they quickly rejected. We have set as a scoring criterion that a subject must entertain an incorrect name as possibly correct for at least 10 seconds in order for the name to be scored as a fabrication. For one of the subjects, S1, a far more detailed analysis has been completed. The details of the analysis are presented in Williams and Santos-Williams (1980). In essence, the analysis classifies each of over 14,000 statements made by S1 in 10 hours of thinking aloud. The classification covers some 30 to 40 categories such as the recall of a full name, the recall of a location, metarnemory comments on the accuracy of information recalled, and strategy comments. The result is a finegrained quantitative analysis of S1 such that we can formally identify complex patterns of events.
...
A Comment on the Validity of Protocol Observations Psychologists have shown substantial concern over the validity of observations derived from protocols (Nisbett & Wilson, 1977). However, we hold that the observations we have derived from subjects' protocols are no less valid than traditional counts of the number of correct items recalled in verbal learning experiments. Because of the very nature of recall, if a subject claims to have recalled that a particular name began with the letter "P," then he must have recalled the information in order to report it. That the subject recalled that the name began with the letter "P" is no more disputable than if a subject claims to have recalled the word "house" from a recall list. Each of the recurrent phenomena we will discuss is defined in terms of either a specific category of information recalled or an ordered set of categories. A second common set of criticisms of verbal protocols is that "thinking aloud" interferes with the process of recall and introduces unknown demand
94
WILLIAMS AND HOLLAN
characteristics. We a g k with both those criticisms. However, the very same criticisms c v be leveled against any psychological experiment. It is simply that so accustomed to ce~tain we have
intermediate steps in the search process. BASIC RESULTS
Some of the basic results are plotted as cumulative graphs against time in Figures 2a-d for Subjects 1-4, respectively. Dotted lines mark the beginning-andend of sessions. The jumps in portions of the curves at session boundaries are due to names subjects reported having recalled between sessions.
NAMES
TIME (hours) Figure 20. Cumulative graph af correct names and fabrications recalled by S1 during ten sessions.
PROCESS OF RETRIEVAL
I
-
m
100
I
-
I I
80
60
0
1
2 TIME (hours)
3
4
Figure 2b. Cumulative recalls by 52. Dotted lines indicate session boundaries.
Even a brief glance at these graphs tells us two things. One, subjects can go on recalling new names for extended periods of time. They were still recalling correct names in the last session after as much as 10 hours (spread over two weeks). Two, the number of retrievable names and fabrications can be extremely large. The number of names ranged from 214 correct names for S l to 83 for S4. In terms of the percentage of possible correct names for each subject, scores ranged from 35 percent for S1 to 12 percent for S3. This seems to conflict with the results obtained by Bahrick et al. (1975). who reported that subjects recalled a relatively small number of names (approximately 40). However, in their procedure subjects recalled names for just eight minutes. Their pilot studies showed that subjects recalled few names after the initial eight minutes, and often reported that they could recall no more. Our subjects gave the same report after a few minutes of retrieval. We simply insisted they continue and they did. In addition to the names recalled, there were literally thousands of pieces of other information retrieved. The information that can be categorized strictly as recalls included: names, partial names, locations, activities, physical characteristics, and personal and social information. Subjects frequently recalled fragments _ C 1 1 of the hornation they were seelung. For example, subjects recalled that a "D" and that it was a two-syllable name.
96
WILLIAMS AND HOUAN
TlME (hours) Figure 2c. Cumulat'm number of names recalled by 53.
TlME (hours) Figure 2d. Cumulative number of names recalled by S4.
PROCESS OF RETRIEVAL
97
Indeed, a great deal of the process of recalling a classmate's name can be typified as a reconstruction from a variety of such bits and pieces of information, This phenomenon has been alluded to in a number of experiments as partial recall (see Bernback, 1970; Woodworth, 1938, p. 37), and is particularly obvious here because we have a precise specification of the target information (names). Sl's finegrain analysis showed over 1000 partial recalls 'and over 10,000 incidental recalls. The protocols also contained strategy and metamemory comments. Strategy comments involved subjects attempting to describe what they think they are doing or are planning to do. Metarnemory comments were characterized by assessments of the validity of the information recalled or judgments of whether the information had been recalled before.
A THEORETICAL PERSPECTIVE
The protocols we have collected consist of thousands of statements. We have identified over 30 distinct categories of these statements. The potential sequences of statement categories one can observe number in the millions. To gather any meaning from data of this complexity it is necessary to have an a priori theory to guide the identification of significant phenomena. The following section introduces our basic theoretical perspective, specifies the kind of evidence of importance to this point of view, and indicates that evidence as it occurs in our protocols. We begin with what we believe are three primary constraints on the retrieval process: partial information, descriptions, and large memory capacity.
Partiai Informution. We believe that a person encodes only a limited amount of the possible information present in the environment. One can think of this limited amount of information, for example, as a list of features or properties, or as partial images. The important point is that not everything that is present gets represented in memory. In effect, information stored in memory can be thought of as a collection of records or packets of limited capacity. Thus, a record might contain a person's last name and not his first. Descriptions. To encode or retrieve any packet of information from memory a partial description is formed that provides an initial entry point into the memory. This notion of descriptions is taken from Bobrow and Norman (1975) and Williams and Norman (1975) (also see the discussion of descriptions in memory retrieval by Norman & Bobrow (1979)). A description is essentially an indexing mechanism. For example, the description that a name is Gennansounding and begins with an "H" could be used as an index for the name Hofstadter.
98
WILLIAMS AND HOLLAN
Large Memory Capacity. We assume that human memory can be treated as if it were indefmitely large, so large that there is always room for more information. We wish to avoid any inclination toward thinking in terms of "economy of storage" or other similar notions. We use the notions of partial information, descriptions, and large memory capacity to define the primitive process for our view of retrieval. We refer to this primitive as the kernel retrieval process. This process takes a small amount of descriptive information as input and outputs a packet containing a limited amount of related information. We propose that there are a very large number of these packets in memory and the crucial problem to be addressed is how one might use this kernel process effectively when engaged in real-world tasks. The kernel retrieval process is diagramed in Figure 3.
IKERNEL RETRIEVAL
I
PROCESS
~igure-3.Kernd retrieval'process.
Such a basic process could be implemented in a variety of ways. For example, it could be simulated on a computer using hash coding, associative memory techniques, or spreading activation. Inasmuch as the input-output characteristics would be the same, we see no need to make any commitment to a particular implementation. We do not intend to decompose this kernel process. In fact, we think that a boundary condition such as this is important to the construction of a unified theory of remembering. We want to use this boundary condition
PROCESS OF RETRIEVAL
99
to put the focus on a higher level of analysis. It is our view that many of the characteristics of remembering derive from the complexity of the world, the large amount of information with which humans must deal, and the flexibilitjl of human remembering, rather than from the nature of atomic retrieval processes. We are imposing this boundary condition to simplify our analysis problem and because of the belief that the details of the implementation of the primitive retrieval process will have only incidental impact on the most salient characteristics of human memory. Just as most of the behavior of gases has to do with the interaction of a large number of particles, we suggest that the global characteristics of remembering derive from the nature of the interaction of a number of elementary processes. We contend that these elementary processes could be implemented in a variety of ways without substantially changing the global ramifications of their intera~on.Much of the behavior of gases, for example, can be understood based upon an idealized model of the interaction between molecules (i.e., elastic collisions). It is the statistical mechanics of these simple interactions that results in most of the salient thermodynamics of gases. In a similar way, we want to explore the global characteristics of remembering by assuming an idealized elementary process and investigating its global implications. The global act of retrieval (i.e., what we observe in human behavior) can be characterized as a reconstructive process based upon the kernel retrieval process we have introduced. The initial information in a query is used to f o descnition of some a s y t of the target i t e ~ T h ekernel process takes this description as input. The information output by the process is used, in conjunction with what is already known, to form a new description. A succession of kernel retrievals leads to a reconstruction of a representation of the target item. We envision that the encoding process is similar to the kernel retrieval process. Because of the same limitation of partial information both in indexing and in the amount of information stored, this process selects some information which is likely to be available at the time of retrieval and uses this as an index to pack away some useful fragment. We do not mean to say that encoding is always optimal or even conscious. Rather, we suspect that humans have developed a set of tacit heuristics for selecting the information to be indexed and the information to stored. Some of these ideas are well known. Bartlett (1932) has presented a view of remembering as a reconstructive process. Dunker (1945), in looking at the nature of problem solving, has considered retrieval as one type of problem solving. His notion of "resonance" is closely related to our idea of reconstruction. Norman and Bobrow's (1979) ideas on descriptions are developed, in part, problem of verifying a collection of retrieval information. Shiffrin (1970) has also examined some of these ideas and his work has influenced our characterization. Norman and Bobrow's (1978) ideas on descriptions are developed, in part, from consideration of our data on a very long-term retrieval.
~
100
WILLIAMS AND HOUAN
The Jigsaw Puzzle Metaphor Imagine the problem of retrieval as being analogous to the reconstruction of a jigsaw puzzle when presented with the pieces from several similar puzzles mixed together. The information provided in the basic query specifies a numbex of starter pieces. One might begin by restricting his view to some likely section of the puzzle, possibly where he has several related pieces (locating this region might be thought of as analogous to finding a context). Next, one searches for a likely piece by looking for something that will match a side with a piece in the "context. " One rummages about in the search domain picking out a piece that fits the description constructed (i.e., has a side which will' complement side of a piece from the context). If it is the case that too mahy pieces match the description, you might want to look for a bordering piece to get further constraints on the piece you are seeking, or you might want to try a set of possible pieces (these actions might be thought of as conducting a search). Finally, you might try to see if the selected piece fits neighboring pieces or attempt to use the newly recovered piece to locate new pieces (this is analogous to conducting a verification). The kernel retrieval process runs into three basic problems which have implications for the global act of retrieval: Too little information, too much information, andfalse recovery.
Too Little Information If there is less information at the time of retrieval than was used at the time of storage, then the missing information must either be retrieved in some new retrieval cycle or it must be inferred. This is the problem of too little information. Imagine trying to retrieve a piece of the jigsaw puzzle which was indexed (encoded) using two neighbors when you have only one of its neighbors. This leads us to expect that subjects would have to conduct incidental or tangential retrievals, guess, or infer missing pieces of information to fill the gaps. This expectation has led to the identification of a number of recurrent phenomena in the subjects ' protocols.
Extended Retrieval. Subjects frequently encountered a situation where a person was identified but not completely named, e.g., "I remember Bill somebody . When this happened the subjects often recalled additional details about the person. We call this extended retrieval. For example:
. ."
S: Bob Peterson, uhh Jerry Paulson. Uhh, this is a rock band. E: What are you thinking about now? S: I'm thinking about a time when they were playing over at o w house. Umm, then's uh E: What isS: I'm'try. I'm-Okay I was imagining the whole room and we w e n imagining the
...
PROCESS OF RETRIEVAL
.
instruments set up and l%tryingg to remember the name of this guy-who used to . . he would also bring a whole lot of people to-fmt on that-what's his name now? Let's s 6 w h i s t t e ) I'm hying toremember his name. At his house was the f m t time I heard a Jefferson Airplane album. Umm .the bass guitar, really a strungout looking dude uhh . . . wow.
Art!And he was also in our 10th grade art class
..
...
...
Frequently such an attempt at extended retrieval would result in the successful recovery of a name. Wih the finegrain analysis of S1 we were able to construct a formal definition of an extended retrieval and thus.count the number of such events (Williams & Santos-Williams, - 1980). There were, for example, 85 extended retrievals in the first 7 of S 1's sessions.
Systematic Hypothesizing. As the difficulty of recall increased subjects began to demonstrate instances of systematic hypothesizing. They generated a sequence of hypotheses, attempting to use the hypothesized information to find additional information. For example, "His first name began with a 'D. ' Don, Dave, Douglas, . . ." Such hypothesizing seems to be a valuable tactic when the number of possibilities is relatively small and readily generated. In comparison, the tactic of extended retrieval seems more widely applicable. Inferential Recalls. In many cases subjects could not explicitly recall a particular piece of information. In such cases, they often confronted the problem of too little information by attempting to infer the information missing. We refer to this as inferential recall. The following is an example: Somebody, . . . I remember this girl who used to play the oboe, and it was junior year, she was our age. Or was she older? I can't remember if she was older or not. I'm sure she . . . now I'm going to check to see if her face was in the yearbook that was ahead of us, the class of 1971. It seems to me as though she might have been next to the one that the year I graduated from. That I was in. I can't remember. The different things that we wore, starting when we were, once again this goes back to the yearbook. When we graduated, these girls wore the scalloped tops. You know, black. And all the years before that they wore a V-neck sort of top (when we went to the studio for graduation pictures). And I'm trying to remember what she wore. Was it scalhped or V ? And that would determine what yearbook she was in. I ' m pretty sure it was scalloped.
The basis for this particular inference was wrong. Scalloped drapes were worn by the female graduates of the subject's high school the year before she graduated. All three of these phenomena, extended retrieval, systematic hypothesizing, and inferential recall fit naturally into a global retrieval process confronting the problem of too little information.
Too Much Information Not all the information available in the query is necessarily useful in the retrieval specification (the description used as an index to locate an information packet).
102
WllLlAMS AND HOLLAN
We refer to this as the problem of too much information. This problem occurs because the encoding process uses only a limited amount of the information to formulate an index. ~uccessfulretrieval requires the use of exactly the inforrnation used as the initial index. There are two distinct aspects to the problem of'too much information which we refer to as the issue of focus and the issue of selection.',The problem of focus is to choose indexing information which is relevant to the target item. In order to search for an appropriate piece of the jigsaw puzzle one must fmt focus down on a likely region. Information from distant pieces must be filtered out. The problem of selection is that less than all of the relevant information may have been used to create the index. Thus, during encoding you may have suspected that only one neighbor of a particular jigsaw puzzle piece is likely to be known when you want to fmd it. If that is all the information used in the index, then if you happen to know both the anticipated neighbor and one neighbor not anticipated, then you need a scheme for filtering out the unanticipated neighbor. Were you to look under an index comprised of information from both, you would find nothing (or a piece from another puzzle). It is important to realize that the problem of too much information does not preclude the problem of too little information. Too much information may be present in the original query, such that some mechanism needs to be imposed to focus down on a likely set of pieces to serve as descriptions to the kernel retrieval process. Even so, too little information may be available in the region of focus, thus requiring the tactics for coping we observed above. Contexts. The use of search contexts by subjects bears directly upon the problem of too much information and particularly on the issue of focus. Subjects made extensive use of locations or activities as contexts within which to search for acquaintances. Names were recovered as subjects thought about "who was on the football team," or "about the time [the members of a rock band] were playing over at our house. " The reference to the P.E. class from the protocol in our introduction is typical. From the point of view of the problem of too much information, a context is a procedure for focusing on a specific subset of the properties in the infonnation already retrieved. Thus, if a subject is thinking about her P.E. class, information retrieved earlier about where she worked after high school can be filtered out and ignored. In Sl's finegrain analysis we have distinguished four major categories of contexts used, 341 unique contexts mentioned, and 987 direct references to search contexts. In the protocols, contexts were often embedded within contexts and strung together in logical sequences. For example, a subject might start with "8th grade classes" and then proceed through "Math class," "Gym," "History, " etc. This -- - of search. leads to the phenomenon Search. While we attribute many of the details of the particular search strategies used to demand characteristics of the task we have set our subjects, we
PROCESS OF RETRIEVAL
103
argue that the use of such strategies is a method of coping with the problem of too much information. We have identified five strategies: Activities, Location, Name Generation, Pictures, and General Association. In this section, we describe these strategies and provide evidence to validate their presence. The details of the searches are not a direct part of the basic argument we are making. Therefore, some readers might wish to skip to the section on the problem of false recoveries. Activities. An approach subjects commonly used was the activities strategy. This stragety is initiated by the identification of an activity in which some group of people engaged. The subjects then proceeded to attempt to name all the members of that group. Examples of groups were: a History class, a Baseball team, Cheerleaders, Band, International club, and the "people I played tennis with. " This strategy was used throughout the protocols of all the subjects. After several sessions it became less and less frequent as the subjects ran out of new activities to work on. In S1's finegrain analysis we identified each instance where S1 recalls an activity. A plot of the frequency of recall of this type of information is shown in Figure 4. It clearly delineates the times when S1 was using this retrieval strategy. ACTIVITIES
TIME (hours) Figure 4. Finegrain analysis of S1: Rate of recall of activities.
WILLIAMS AND HOUAN
Locations. The locations strategy involves a subject systematically searching a mental map where target items are likely to be recalled. An example of the use of this strategy is the following: That's a new name. He lived on Alvian Street. Ah, let me see, on the other side of there, there is Bob. He was a year older and that other girl was a yearolder. O.K., so i m there is Margaret Mon and if you come up and around, there is no one right there. J Gon lived down there, but I've already named him. I already named Judy Nicholson and let me see if there is anybody on the cross street. There was a girl that was younger and she had a brother that was a year older than me. Then there was a girl a year younger and there is Gay Masterson, I already named her, and Barbie Tollen. They live on tfie same street and they have no one else on that street that went to Point Lana with us. If I keep going down Silvergate there is someone who lived on the comer house. It had a purple door . . .
A plot of the frequency of S l 's recollections of locations in Figure 5 shows that S1 used this strategy throughout the second and third sessions. Name Generation. The name generation strategy involves a subject inventing a scheme to generate common first or last names and then testing the names to see if they match the names of classmates. One common method was to proceed through the alphabet, generating common first names. For example: LOCATIONS
TIME (hours) Figure 5. Finegrain analysis of Slr Rate of recall of locations.
PROCESS OF RETRIEVAL
105
Are there any other Bettys tbat I knew? I thirrk she's the only Betty I ever knew. Betty, other girls names with "B"s tbat are sort of familiar names. Barbara and, I named Barbara Shafer already, and there was Barbie Tollen and I named her. Barbara, Baxbie, no those are the only two Barbaras I know. Umm. another "A" name for a girl might be, umm, there is Ann.I don't know any Alices. Umm, no Allisons. AM. What other "A " names.
This generation of names from the alphabet also frequently was used to produce last names. Another method that subjects used was to look around the room naming objects and generating names from the objects' names. Although this later form of name generation occurred in short bursts throughout all sessions, systematic searches using the alphabet as an organizing tool generally did not occur until the later sessions. S1 began using the strategy as early as the fourth session; S3 did not use the technique until the tenth session. S l 's use of the alphabet as a generating tool is of particular note. She used the technique extensively for nearly five hours. Initially she covered the entire alphabet in less than one hour. She then made succeeding "deeper" passes using as many as four alphabetically generated letters to start a name. In the eighth session the entire hour was devoted to female names from Be-- to De--. Figure 6 shows S 1's frequency of naming of letters or sets of letters over the ten sessions. Pictures. Another common strategy used by the subjects was what we call the pictures strategy. Subjects would scan some internal "image" of a picture or set of pictures to locate individuals whom they would then attempt to name. Frequently the pictures were out of the subject's yearbook or from a collection which the subject had a strong a priori reason to believe would have classmates' pictures (e.g., a summer camp picnic). The following is sample of this kind of search: O.K. who else? Umm, Ronnie Walker, she was also the last letter. If I see her picture I can-like there was a lineup of girls who were on the intramural volleyball team. Brenda was one. Thae were some younger people too. And Ronnie Walker was one and
. ..
The protucols sound as though the subjects view veridical images. However, there is some evidence to suggest that these images can have interesting flaws in them. For example the following protocol is a sample of picture searching by S3: . . I'm going through the yearbook again. I'm hying to look over the faces and maybe even try to visualize the names next to them, next to the pictures. Maybe that will help. I don't think it will. Sharon Farley. She was pictured next to Mark Farley and I member her. S h a m Farley.
It all sounds very plausible. And indeed, both Sharon Farley and Mark Parley are in S3's class. However, Mark Farley's picis not in the yearbook. Rather, his name is listed among those did not have their pictures taken. The image from which S3 is abstracting her search information is wrong: it never existed in the real world.
WILUAMS AND HOLLAN
LETTER SETS
TIME (hours) Figvre 6. Finegrain anabir
of S1:
Rate of recall of I*
sets.
General Association. Every subject used the strategy of general association at one time or another. Indeed, it appeared that most subjects began their recalls with this strategy. General association is the strategy of starting with a known person and searching for people who are d i i y related in some obvious manner (e.g., a friend, cousin, brother or sister). Occasionally a particular person will suggest a group of people (e.g., a social clique). In the early stages, subjects' protocols were characterized by the dense retrieval of names with infrequent identification of the explicit context. In most cases the use of this strategy is inferred when the subjects report the relationships of a group of names in a post hoc manner. For example: The guy Me1 Hagershon that I had a hard time thinking of his last name. I remember his friend was named Mike and I couldn't nmember his last name. Fielding, or something like that sticks into my mind. Maybe.becausef b o w a Mike Fields that isn't related to school but that name also sticks into our mind for him although it may not be right at all. Ah, I thought there was something else I thought of. Let me see. Is there anything else. Jonny Faylan. He was also friends with that group, with Me1 Hagershon and Mike and I knew him since I was young.
Search strategies in our view result fmm the systemati;c application of search contexts. They are generally characterized by the embedded retrieval of
107
PROCESS OF RETRIEVAL
one context within another. Examples of this embedding can be seen in the activities strategy, the locations strategies, and some name-generating strategies. The recursive application of this embedding creates a tree-like search. Some search strategies can also be characterized by the repeated application of the same context to different objects (e.g., general association). We found that subjects adopted specific search strategies for periods of time from a few seconds to several hours. On occasion, subjects briefly shifted from one strategy to another that looked more productive, returning when it was exhausted. We interpret these search strategies as methods for dealing with the problems of too much information and of too little information. Search strategies provide a systematic mechanism for successive focusing of the retrieval process. The systematic nature of many of the search strategies provides a mechanism that tends to cover most aspects of the existing information.
False Recovery The third problem confronting the kernel retrieval process is false recovery. Because descriptions are used to guide retrieval, a record about an item similar to the one being sought may be recovered. Thus one can imagine retrieving a piece from the wrong jigsaw puzzle simply because it fits the neighbors used to generate the index. We see indications of false recoveries in verification processes and in the errors subjects made. Overshoot. Sometimes when's subject recalled the name of a friend, she continued to recall additional information. For example, "Steve Jones, he was a box boy at the Food Basket [store] on the comer, and was on the football team. " We call this phenomenon overshoot. One can picture verifying a newly recovered jigsaw puzzle piece by seeing if it yields new pieces which fit: Although in some cases overshoot may be an artifact of conversational constraints in the protocol situation, in other cases it is clearly related to a process of verification. Even though a name is recalled, it must still be determined that the name belongs to the person being thought of, and that the person attended the correct high school in the correct year. For example, in the following protocols overshoot is used overtly as an attempt to verify a name (the overshoots are in italics). Okay. Oh, oh, wait a minute. Now we're back to a class I haven't really thought of before, our world affairs class. There was this guy who used to sit in back of me. And he took Spanish classes and I know his name. Hi last name began with an "0." Ah . . . The name Orin Elliot sticks and that is the fmt time I've not associated with someone in that history class. Not really a history class, a political science class. He sat in backof me he had dark hair ah, I think . . He was Chin Elliot. That was his name.
...
.
Or an example where the attempted verification fails: Carol Hammer, I already mentioned her. Oh and Turner. Linda, Linda Turner. I don't know about her frst name, but I know her last name. A blond girl. Yeah,Turner. I
108
WILLIAMS AND HOUAN
don't t h i that's her name. I don't think Linda is her name. I think it's ju&hmer, for sure. She wound up in Redlunds somewhere. I remember that. I don't remember her fmt name. Linda Turner . That seems to fit. I don't know. I don't know for sure, but that's the closest I can come. The last name for sure I remember.
..
Indeed, Linda is not the correct fmt name. As in the case with extended retrievals these. complex events can be fonnally defined and quantified. For example, S1 exhibited 21 1 simple overshoots, and 29 verification overshoots (those which gave evidence of uncertainty about the original information recalled) in the first 'l sessions.
Fabrications. Fabrications are defined as those names which the subject recalled but which are not in the subject's yearbook. We have chosen the term fabrication to suggest some of the reconstructive character of these errors. Fabrications fell into one or more of five distinct categories: acoustic errors, wrong graduating class, misassignment, missing picture, and those of indeterminate origin. Some incorrect names were obvious acoustic errors. For example, "Lloyd Chappin" turns out to be Lloyd Chaffin (this was confirmed during the subject debriefing). Some fabrications can be identified as names of students who did not graduate with the subject's class but were in the subject's high school in a different class. In other cases the subject assigns one person the name of another individual (this event can be detected only in unusual circumstances, for example, when the subjects catch themselves in an errgor in followup experiments when pictures were used). Some names recalled by our subjects were those of people who were in the subject's grade in earlier years but for whom there are no records in the 12th grade. We are not certain whether these people moved away, did not graduate, or simply missed having their pictures taken (this category is relevant only for subject S 1; yearbooks for the other three subjects had lists of all those who did not get their senior pictures taken). The final category is fabrications of indeterminate origin. These are names which we could not independently track down. They may be people who were friends of the subject but who went to different schools. Table 1 shows the number of fabrications in each category for each subject. Frequently, fabrications were events of substantial structure. Many were caught and corrected by the subject, some shortly after the error was made, some a number of days later. Some fabrications appeared to be simple recalls; others were accompanied by extended retrievals andlor obvious uncertainty about the recall (as was the case with many correct recalls). Some fabrications were highly suggestive for the underlying processes which led to an error. For example, the name Carol Ludlow was initially recalled but was corrected several days later to be Judy Wardlow. Karen Ludwig is a friend of Judy Wardlow's and was often recalled as "Karen Ludlow . No! Ludwig. Karen Ludwig. " The fabrication "Carol Ludlow" appears to be a confounding of the names Karen Ludwig and Judy Wardlow.
. .
PROCESS OF RETRIEVAL
TABLE 1 Number of Fabrications per Category Subjeds Categov
S1
Acoustic Similarity Wrong Ywr Misassignment of Name Missing Piclure?* UnknodOther Total**
35 8 1 26 38 100
S2
S3
5 17 10 26 0 1 2 O* 0' 11 25 25 69
S4 25 12 3 O* 56 90
*Yearbooks for S2, S3, and S4 listed all A o r s who did not have pichres taken. "*A few fabricako~fell into more than one category.
A few instances show the difficulty of r&onstructing a name from partial information. For example, one subject recalled someone named "Bill . . or maybe Bob . . . [who had a last name] like Proxmite, but it wasn't Proxmire." The correct name is Bob Billmire. All of the elements of the name are accessible, but the subject put them together in the wrong order. "Bill" is the fmt syllable of the last name, not the first name. That some of these fabrications appear to be from the incorporation of incorrectly recalled fragments of information points directIy at the problem of false recovery and is consistent with our assumption about partial information. The flaws shown in the acoustic fabrications seem clearly indicative of the reconstructive natuxe of recall.
.
Self-Correction. On occasion subjects changed their minds. A name was recalled only later to be corrected. For example, "Fred Knight, no . McKnight! That's it." We call this self-correction. This occurs spontaneously without any information or hints provided by the observer. The very existence of s e l f - c o d o n s means that at least one verification process must exist. How else could subjects change their minds after a retrieval has been conipleted?
..
DISCUSSION
The only active characterization of the global process of remembering we have given is the jigsaw puzzle metaphor. The metaphor is truly a metaphor and not a process model, and furthermore, it is seriously flawed. We are not far enough along in our research to produce a true process model. Yet we find ourselves in need of some sort of active characterization of the collection of tactics and techniques we have observed to summarize our view and serve as the basis for a discussion of some traditional memory phenomena. Therefore this section begins with a description of what we call the retrieval cycle and follows with our view of
110
WILLIAMS AND HOLLAN
forgetting, retrieval errors, search, and the distinction between recall a r ~ccog~ ~ nition.
The Retrieval Cycle From our observations the retrieval cycle seems to consist of three phases (shown in Figure 7), each intended to confront one of the basic retrieval problems. Fist, there is the establishment of a retrieval context. This is designed to focus on a relevant subset of the initial information provided. Next, there is a search through the memory space reachable from the context and the available information. Third, there is verification of the information retrieved. If the information retrieved satisfies the original query, the retrieval terminates at this point. Otherwise, the retrieved information is used to reformulate the description and a new cycle is initiated. This new cycle may or may not require the establishment of a new context.
FIND A CONTEXT
Figure 7. Three phases of (he retrieval cycle.
PROCESS OF RETRIEVAL
111
One important aspect of the retrieval cycle is that each major stage.is recursive. It is this recursion that we find so difficult to capture in the jigsaw puzzle metaphor; it has led us to create this alternative description of the global retrieval process. This recursion expresses itself in a host of ways. The establishing of a context ("the volley tennis group in P.E."), for example, may itself require a retrieval cycle, involving the finding of a context, search, and verification. The embedded nature of some of the search strategies we have identified provides a direct indication of recursion. Once a context is established, the search can begin, but it too may require one or more retrieval cycles (e.g., an extended retrieval), this time contained within the higher-level search context. Finally, the verification phase requires its own retrieval cycles for the purpose of certifying the accuracy of the information provided by the preceding search phase. In the case of overshoot, for example, successful additional retrievals are used as a validation. Retrieval of a Context. The fist subprocess of the retrieval cycle isfind a context. The context simplifies the problem of forming a description. However, the recovery of the appropriate context is itself a retrieval problem. If that is so, how can the retrieval of a context help? Consider why we might need a context from the point of view of the encoding of information. At the time of encoding, how does one choose a set of properties to use as a description? There are two competing factors. The properties should produce a description of a unique record. Otherwise, with a limited description, we might expect to have to search through a large collection of records to find the right one. Second, the properties must be such ,that it is possible to generate them at the time'of retrieval. The first factor suggests that we should use as many and as diverse a set of properties as possible. The second factor suggests that we should use some limited and stereotyped set of properties for encoding. The notion of a context seems a compromise between these two extremes. The key leverage the notion of a context provides is that it can be recovered with relatively general information, can guide and focus the search for more specific information. Thus, you need only know about high schools in general to pull out the idea of team sports, from that you can get more specific information about football teams, from that comes information about your particular high school's football team, say that they practiced in the old field behind the gym, from that you might be led to a particular incident of a practice you once watched, from which you might identify some individuals to name. Search Within a Context. The second stage of the retrieval process is search within a context. Given a context within which to work, the retrieval cues provided in a question may be insufficient to uniquely specify the record being sought. In such a case, a description-based system has two alternatives. The first
112
WILLIAMS AND HOLLAN
alternative is to recover records about the item (from some other context) for which a sufficient set of retrieval cues does exist. In this manner more information about the item is retrieved from which still more information can be gathered until enough retrieval cues are recovered to generate a good retrieval specification. This is exactly what we observe in extended retrieval. The second altemative is to check out all of the records which can be retrieved with the limited set of cues available by hypothesizing additional cues. This we see in systematic hypothesizing.
V e r i f ~ c ~ o The n . final stage of the retrieval process is vertfi. When a record is recovered in the search process there is some chance that it will be a false recovery. The verification process is a natural consequence of this problem and the necessity to confirm inferential recalls. The verification process is intended to identify false recoveries and to reduce the possibility of being misled by a distracting inferential recall. In the retrieval system we propose, there are several ways information can be verified. One technique is coincident recovery. Once a piece of information has been recovered it can be considered as simply a possibility until it has been confirmed by an independent recovery of the same information. A second verification technique is indirect confirmation. If a subject can use a bit of information just recovered as part of a description to retrieve additional information, then the subject could judge the information to be verified. This technique may be the source of the overshoots observed. A third possible verification technique is consistency checking. If the information recovered fits with what is already known then a subject could consider the information likely to be correct. Thus, if a subject recalls that a particular person, who is known to be on the football team, is big and heavy set, then this new information is consistent with what is already known about football teams in general. Isolating verification techniques in this manner makes it possible to think of the verification process as a set of subprocesses, each of which filters out some percentage of errors. These filters also might eliminate good information along with the bad. In practice it is probably the case that all of these techniques (and others not considered) are used together in various combinations during any specific recall to reduce, but not eliminate, retrieval errors. Another way of thinking about the verification process is as an integration of the application of the techniques mentioned. Thus a subject might judge the "connectedness" or consistency of a large array of fragments of information recovered about a particular.person. Judgments of how tightly bound any piece of information is to the whole (e.g., Is it consistent with what is known? Can it be used to recover other information which is known? Can it be recovered in more than one way?) may serve as the basis for judgments of certainty. Though this second way of conceiving of the verification process is more complicated, it is the view we favor. We propose that lhe degree of certainty that subjects express
113
PROCESS OF RETRIEVAL
about the validity of their reco'llection is based upon the extent of success they have had in verifying it.
The Interpretation of Some Memory Phenomena The foregoing characterizationof remembering provides a framework in which a number of traditional memory phenomena can be interpreted.
Forgetn'ng. One of the most salient phenomena of human memory is forgetting. Frequently, memory models account for this phenomenon by introducing notions of decay or unlearning. The characterization of memory we'have introduced has no provisions for the decay or deletion of information once it has been stored. How then is it possible for this characterization to account for forgetting? We believe that forgetting is caused by four interacting retrieval problems which confront the global retrieval process: false recovery, reencoding, too little information, and too much infonnation. As a result of these problems and their interaction a variety of conditions can make a specific piece of information more difficult to recover. If the difficulty becomes too great, the search is abandoned and the infonnation is reported to be "forgotten. " Notions similar to this have been proposed before in the literature. The distinction between availability and accessibility (Tulving & Pearlstone, 1966), is an example of such an idea. One important difference is that what we are proposing is a mechanistic account for such retrieval failures. Any time an event is being reconstructed (i.e., recalled), it is possible to recover a record about a similar event and incorporate it into the reconstruction of the event being recalled. When this "merging" of two events occurs a new event has been constructed. Because the retrieval process is using infonnation about the event to find contexts, to search within contexts, and to verify the recoveries, the retrieval of new information can become blocked. The fabricated event never actually occurred, therefore no information about it exists. Within the jigsaw puzzle metaphor one can envision the incorporation into the puzzle being built of a piece from some other puzzle. The piece fits locally. The problem,is that it interferes with the rest of the reconstruction. It may suggest inappropriate contexts, it may exclude the incorporation of a correct piece, it may interfere with the validation of another piece. The following protocol from S 1 is an example of a search being misdirected by a false recovery.
.
Nancy Phillipano, . . she had a sister that was a y+ younger, but also graduated with our class. and her name was. Because she would have been in our senior class too. Last name was Phillipano. There was Nancy Phillipano and her sister, who was . Her name began with an "R." Rochelle, or, umm, not Rochelle. Her sister was sort of pretty! and she always wore it, it was long and sort of real real curly, and she'd always wear it like with the top part pulled back into a pony tail in the back and would let the back hang down. She would always wear a suedejacket. And what was her name? Seems to me that it began with an "R," umm me see Not Robbie, not
..
...
114
WILLIAMS AND HOLLAN
.
Rachel, umm . .Phillipano, Nancy Phillipano. And what was her sister's name? Not Rachel, umm . . .Not Ramona not umm, let me see. O.K. I'mtrying to think of "R" names. She had a sort of unusual, not Rachel, not Roberta, not Robin, Robin Phillipano? No. Roberta Phillipano, umm . . . Let me see. Umm, let me see, umm, Not Rolly, umm. Can't think of any more "R" names maybe (mumble). With an "R." But she did graduate with our class. She graduated a year early. Her last name was Phillipano. Who did she hang around with? Let me see, of other people who are (mumble). Diane Hart! That's a girl I couldn't think of her name. That's one girl's name. That's a new name.
In another session S 1 recalls (correctly) that Nancy Phillipano's sister is named Lou AM. AS long as S1 was looking for a name which began with an "R," Lou AM could not be recovered. Among other things, the problem of false recovery suggests that the more events of a similar nature encoded in memory, the more difficult the task of retrieval becomes. Thus, response competition, retro-pro-active inhibition, and other like phenomena can be interpreted as a consequence of this mechanism for retrieval failure. Any time we recall information about an event we have an opportunity to reencode it. These reencodings may be different from the original encodings in a number of ways. Often, reencodings are abbreviations of the original encoding. For example, only some limited amount of information about the item may be needed to answer a given query. Therefore, the retrieval cycle can quit once the critical piece(s) of information has been recovered. Also, some of the information may be "forgotten." Reencodings can contain new information which may have been inferred or deduced during the recall. This may take the form of filling in bits and pieces of information about the event or the information recovered can be reprocessed to discover new things. Reencodings may also contain incorrect information coming from false recoveries built into the reconstruction. Retrieval after one or more reencodings can be thought of as the reconstruction of a given jigsaw puzzle after pieces from several puzzles about the same topic have been thrown into the pile. Many of these new pieces may facilitate the reconstruction, but in certain circumstances, they can also interfere. False recoveries put into the reencodings (e.g., Nancy Phillipano's sister's fist name begins with an 'R') clearly can interfere. Correct information added in new encodings can suggest inappropriate contexts (the problem of focus) or may overspecify the retrieval index. Sketchy reencodings tuned to retrieve information from one query may guide the retrieval away from information sought in some new query. As we pile reencodings atop reencoding the original event tends to become stereotyped (inferences), abbreviated (less information), tuned toward more common queries, or even wrong (false recoveries). Such a course of events might well lead to the distinction between semantic and episodic memory that Tulving ( 1972) has proposed. Bartlett 's ( 1932) observation that repeated recalls become successively more stereotyped also fits well with this mechanism. In this light, it is interesting to note that mnemonists report that when
PROCESS OF RETRIEVAL
115
information is encoded using a mnemonic strategy, after the information has been recovered a number of times using the scheme, the mnemonic seems to disappear and all that is left is the information which was originally encoded (see Lorayne & Lucas, 1974). In the case of too little information the retrieval process confronts three problems. One, a false recovery is much more likely because little information is available to use in the verification process and because the number of possible events described by the information is large. Two, it becomes xi-iore likely to recall a reencoding of the event rather than the originalevent. This occurs, because sketchy reencodings may be encoded using fewer and perhaps more typical properties. Three, if a property which was used for indexing part of the knowledge about an event is not accessible from the information presented in the question, then the knowledge indexed under that property can be recovered only by guessing. If a few such properties are not available, the guessing combinations become numerous. In the case of too much information the retrieval problem is essentially one of trying to determine the correct context to use. If the subject is misled into choosing the wrong context, the retrieval task becomes impossible. The phenomenon of encoding specificity (Tulving & Thomson, 1973) appears to be just such a situation. It is not the case that we view the above as fo& hdependent mechanisms for forgetting. Rather, all of the retrieval problems talked about in this section interact with one another. The problems of too little information and too much information may occur on different cycles of the same retrieval task. Subjects may begin a task with too much information. Once they find the correct context, they may not have enough information within that context. Too little information is a problem, in part, because of the increased probability of a false recovery, or of recalling a reencoding of the item sought. There is no single simple cause of forgetting. There are a variety of major influences on the process of retrieval which lead to forgetting.
Search. The phenomenon of search as a natural product of our notion of a recursive retrieval cycle has been extensively explicated in earlier sections. Any difficult retrieval task, almost by definition, results in a search. What is surprising, from the point of view of our characterization of retrieval, is the occasional "instantaneous" recollection. Sometimes a recall appears to take no search at all. We suggest that even these rapid retrievals require some search. The difference lies in the length and the complexity of the searches. The common phenomenon of search is the result of those retrievals which take long enough for the intermediate steps of the retrieval to be distinguished. Even in recalls that appear to be instantaneous, information beyond what is sought frequently comes to mind. Whether this other information occurred before or after the retrieval of the target information is uncertain. Our view is that all recollection can be placed
116
WILLIAMS AND HOLLAN
along a dimension of the amount of search. In long investigations the search processes can be easily identified, while recollections which appear to be instantaneous are simply brief searches. A variety of phenomena can be interpreted as evidence for this view. In brief: 1. The continuity of recognition reaction times. If we delay the recognition process, the time required to recognize an item increases. Thus, if we test subjects immediately after the presentation of a list, they will recognize items faster than if we test them after a delay (Woodworth, 1938, pg. 38). This increase in reaction times can span from immediate recognitions with no apparent search to lengthy recalls in which search protocols can be obtained. There is no sharp increase in reaction times which might suggest distinct processes. [There is some evidence contrary to this in the hazard functions reported by Fumas, 1980.1 2. Intrusions. Many of the errors which occur in instantaneous recalls have the same reconstructive character as those errors obtained from lengthy searches. A prime example is speech errors. Although retrieval time per word is very brief, speech errors often involve an item which is similar to the target itme, or the improper reconstruction of the correct pieces (Baar & MacKay, 1978). 3. Retrospective reporting. Occasionally, subjects can report, in retrospect, an involved sequence of retrieval steps which led to a recall occurring in a brief instant. For example, the following is a segment of Sl 's protocol during an extended period of using a location search strategy: Okay, let me see. Does anybody else live back in there that we h o w now: Oh! Maxine Levin. She didn't live there, but Pam Volpone now lives on Loma Portal and Mike Stevenson, who is eleventh grade, used to live in the house that Pam Volpone and her husband bought. And so from Pam Volpone I got Maxine Levin 'cause I saw her at the store the other day and she was talking about Pam Volpone.
This protocol suggests that the subject first recovered a house in which a friend who was a year younger, Mike Stevenson, lived; and from that recalled that a classmate, Pam Volpone [named earlier], who was now living in the house; and from that recalled that the first classmate, Pam Volpone, had been the recent subject of a discussion between the subject and another classmate, Maxine Levin. If the subject's report is correct, then all of these recollections occurred within the fraction of a second pause between sentences.
Recall and Recognition. In the characterization of remembering presented in this paper, the common distinction between recall and recognition fades into the dimension of search. In the classic recall paradigm: the subject is given a minimum of cues from which to begin a retrieval; in the recognition paradigm the subject is given an abundance of cues from which to initiate the retrieval. The result in the first condition is a comparatively difficult and lengthy search, whereas in the second condition the search is generally easy and therefore brief. Mandler (1976) points out the lack of a sharp distinction between recall and
PROCESS OF RETRIEVAL
117
recognition and the common confusion between the theoretical notions, the experimental paradigms, and the colloquial meanings of the two terms. In our view that it is not useful to make a sharp distinction between recall and recognition. Rather, one should think of the two notions as two poles along a dimension of density of useful retrieval information. As the amount of information presented increases, remembering becomes more like recognition and less like recall. The problem in most recall paradigms is too little information (resulting in the-search for more information); the problem in most recognition paradigms is too much information (resulting in the need to choose the appropriate context).
Intrusions. Intrusions are errors in the recollection of an item. Frequently they are considered to be any kind of erroneous word recalled in a standard word learning experiment. Sometimes the term can refer to the recollection of an event that is the confounding of two or more separate events. Intrusions have also been referred to as fabrications or simply errors. The characterization of the retrieval process which we have developed here has as logical consequences not only the simple fact of intrustions but also many of their observable characteristics. Intrusions come both from false recoveries and from inferential recallsgecause only partial information can be used in the specification of a record, the record which is recovered may be the one originally encoded about the event being sought or it may be a record about a similar yet distinct e v e n t 3 s false recovery may contain information which is not true with respect to the event being sought. These false recoveries have several results: 1. The false recovery can be incorporated in the reconstruction of the information being
output. For example, the acoustic errors which were observed as fabrications. 2. The false recovery can misdirect the search for more information. This misdirection can itself have several results. 2a. The false recovery can cause the search to fail. The information recovered may lead away from the correct records. It may also yield faulty verifications, resulting in rejection of appropriate records. The section of Sl 's protocol presented in which she is searching for names that begin with 'R' is a good example of how a false recovery can misdirect fhe search. 2b. The search may be directed to additional false recoveries, resulting in the fabrication of an event which never actually occurred, or in the recovery of an item similar to the target item. For example, the misassignment of names which were observed as fabrications.
Inferential recalls could be considered intentional false recoveries. Inferential recalls are recollections about items similar to the one sought or about the
general class of items of which the target item is a member. Thus, in recalling that a person's name is "German sounding," the subject goes on to recall information about "German-sounding names" to be used in guiding the search.
118
WlLUAMS AND HOllAN
Each "German-sounding name" that is recalled in this fashion is, from this point of view, an intrusion. Inferential recalls can have the same effects on the retrieval process as false recoveries. Thus, an inference can be incorporated into the reconstruction and occasionally misdirect the search. Note that inferential recalls :need not necessarily misdirect the search. Indeed, their main function is to provide intelligent guidance to the search process. One of the primary focuses of our planned research will be on an account of the exact errors people make in recall. We believe that it is in providing such an account that theories of reconstructive recall hold their greatest promise as well as their greatest challenge. We think it is important to examine carefully not just the number of errors subjects make but also to be able to giveL detailed descrip--.... _ tions of their nature and accounts of their origin. CONCLUSION
We are taking seriously the notion of remembering as a problem-solving process. To do so we have defined an idealized kernel retrieval process as a boundary for our analysis. Given this kernel retrieval process, we have focused on the interactive nature of the global process of retrieval to build a characterization of remembering. From the three basic constraints which we have called partial information, descriptions, and large memory capacity, we have suggested a three-stage characterization of a reconstructive retrieval process. Information about the target item is used to construct a description of some aspect of the item. This description is used to recover a fragment of information about the item which is added to what is known. From this information a new description is formed to retrieve still more information, until the particular piece of information sought can be recovered. The three retrieval stages are:find a context in which a proper environment for conducting a search is recovered; search in which bits and pieces of information appropriate to the context are recovered until an adequate description can be formed within the search context; and verifj, in which the information recovered is checked for consistency. Each of the three stages can have within it one or more recursive calls to the retrieval process. This view of remembering can provide an account for the types of information subjects retrieve during the conduct of a lengthy intentional recall task. In addition, it results in a new account for many traditional memory phenomena. The account of forgetting, for example, proposes a relatively detailed mechanism for forgetting as a consequence of retrieval failure alone. Forgetting as retrieval failure is not new. A detailed mechanism which distinguishes between forgetting caused by false recovery, too much information, too little information, and the problem of reencoding is new. The characterization we have developed is not a theory of remembering. Rather, it is a set of distinctions which we intend to use to build a theory of
-
PROCESS OF RETRIEVAL
119
remembering. The accounts' we have given to explain the collection of phenomena we have addressed are only outlines of the mechanisms, a set 'of goals we intend our final theory to yield. ACKNOWLEDGMENTS We would especially like to acknowledge support, criticism, and continued interactions with Don Norman and Dave Rumelhart from UCSD, Allen Collins from BBN, and John Brown and the other members of the Cognitive and Instructional Sciences Group at Xerox PARC. This work has received support from a number of sources: Office of Naval Research, 'Navy Personnel Research and Development Center, Advanced Research Projects Agency, and Xerox Corporation. The views expressed are those of the authors and should not be interpreted as representing the official policies of any of these agencies.
REFERENCES Baars, B. J., & MacKay, D. G. Experimentally eliciting phonetic and sentential speech errors: methods, implications, and work in progress. Language in Sociery, 7 , 1978, 105-109. Bahrick, H. P., Bahrick, P. O., & Wittlinger, R. P. Fifty years of memory for names and faces: A cross-sectional approach. Journal of Experimental Psychology: General, 1975,104. No. 1, 54-75. Bartlett, F. C. Remembering. Cambridge: Cambridge University Press, 1932. Bembach, H. A. A multiple-copy model for postperceptual memory. In D. A. Norman (Ed.), Models of memory. New York: Academic Press, 1970. Bobrow, D. G., & Norman, D. A. Some principles of memory schemata. In D. G. Bobrow & A. M. Colliis (Eds.), Representation and understanding:Studies in Cognitive Science. New York: Academic Ress, 1975. ' Dunker, K.On problem-solving. Psychological Monographs, 1945,58, 1-1 13. Fumas. G. Personal communication (1980). Lorayne, H., & Lucas, J. The memory book. New York: Stein and Day. 1974. Mandler, G. M. Memory research reconsidered: A critical view of traditional methods and distinctions. (CHIP report No. 64), La Jolla: University of California at San Diego, Center for Human Information Processing, 1976. Nisbett, R. E., & Wilson, T. D. Telling more than we can know: verbal reports on mental processes. Psychological Review. 1977.84, 233-259. Norman, D. A. Toward a theory of memory and attention. Psychological Review. 1968, 75, 522536. s : intermediate stage in meyory r&iival..~o~niNorman, D. A., & Bobrow, D. G. ~ e s c h ~ t i o nAn tive Psychology, 1979,II. 107-123. D. A. Norman (Ed.), Models of human memory. New York: Tulving, E. Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press, 1972. Tulving, E.. & Pearlstone, Z. Availability versus accessibility of information in memory for words. Journal of Verbal Learning a d Verbal Behavior. 1966,5. 381-391. Tulving, E., & Thornson, D. M. Encoding specifity and retrieval processes in episodic memory. 1973.80, No. 5 , 352-373. Williams, M. D.. & Norman, D. A. Some characteristics of retrieval from very long-term memory. Paper presented at the meeting of the Psychonomics Society, St. Louis. Mo., 1976. Williams, M. D., & Santos-Williams, S. M. A method for exploring retrieval processes using verbal Nickerson (Ed.), Attention and performance VIII. Hillsdale, N.J.: Lawrence Erlbaum Associates, in press. Woodworth. R. S. Experimental psychology. New York: Holt, 1938.
