(11) PSY452 - Chapter 8: Models of Semantic Memory
11. What is the evidence that inferences based on schemas can sometimes lead us to misremember information? Discuss, citing the classic work of Bartlett (1932).
-Bartlett believed that the most interesting aspect of memory was the complex interaction between the participants' prior knowledge and the material presented during the experiment. -In particular, he argued that an individual's unique interests and personal background often shape the contents of memory. -In Bartlett's (1932) best-known series of studies, he asked British students to read a Native American story called "The War of the Ghosts." -They were then asked to recall the story 15 minutes later. -Bartlett found that the participants tended to omit material that did not make sense from the viewpoint of British students. -For example, they often omitted a portion of the story in which a ghost had attacked someone, and this person did not feel the wound. -These students also tended to shape the story into a more familiar framework so that it made sense from a British perspective. -In many cases, the students' version was more similar to British fairy tales. -Bartlett also asked his participants to recall the story again, after a delay of several days. -As time passed after hearing the original story, the participants borrowed more heavily from their previous knowledge and included less information from the original story. -The research also shows that schemas can mislead us, so that we make systematic errors and "remember" information that was not actually stated. -In our daily lives, however, these schemas are usually helpful, rather than counterproductive. -For instance, our background knowledge can help us recall stories from our own culture. Simple stories have definite, regular structures -We become familiar with the basic structure of stories from experiences in our culture. Then we use this structure to interpret the new stories that we hear. -In general, then, this background information helps us draw correct conclusions.
8. How are boundary extension errors related to schemas?
Boundary extension refers to our tendency to remember having viewed a greater portion of a scene than was actually shown Notice that the earlier topics in this discussion of schemas were verbal; in boundary extension, however, the material is visual. Still, our schemas for complete objects help us fill in missing material during a memory task. -Helene Intraub and her colleagues were the first researchers to document the boundary-extension phenomenon. -Intraub and Berkowits (1996): showed college students a series of photos like the garbage scene in Demonstration 8.6. Each photo was shown briefly, for 15 seconds or less. -Immediately afterward, the students were instructed to draw an exact replica of the original photo. The participants consistently produced a sketch that extended the boundaries beyond the view presented in the original photo. -As a result, they drew more of the background that surrounded the central figure, and they also depicted a complete figure, rather than a partial one.
6. Now that you've described the basics of PDP models, semantic networks, and propositional networks, compare and contrast the network vs. PD approach. Consider how they differ in how knowledge is represented and stored and the sorts of cognitive processes they can model.
Models: the take home message -It appears that memory is organized based on meaning -The idea of spreading activation has helped us think about knowledge -Still debate over whether PDP models are preferable to semantic networks > local vs distributed representation Network vs PD Networks: Info is represented by symbols (nodes) in our mind (local representation) -Nodes, links, networks -Spreading activation -Network models suggest content addressable storage (like us) -Accounts for typicality effects well (lots of evidence from priming studies) -Networks can explain a lot of data and remain very popular PDP: concept representation is distributed across nodes (distributed representation) -The combination of nodes represents the concept -The nodes are sub-symbolic (they stand for symbols) -Meaning is the pattern of activation (scantron) -Can be used to model: (children's language learning, decision making, the Stroop effect, memory deficits in Alzheimer's disease) -Can respond to new things (generalizes) -Mimics humans in many ways (a lesion doesn't totally shut the model down --> graceful degradation)
4. Describe the propositional network approach. In your answer you should: a. Discuss what is meant by a proposition. b. Describe how propositions are represented within these networks. c. Describe how these networks differ from older, semantic networks
Propositional networks: modern version of semantic networks a. Base unit is the proposition: smallest unit of knowledge that is true or false -ex: Dogs chase cats b. ex: Dog is the agent, chase is the relation, cat is the object c. Network models -Suggests content addressable storage (like us) -Networks can explain a lot of data --> remain very popular
7. Is memory better for schema consistent or schema inconsistent material? Discuss, citing a study in support of each idea. Next, discuss the current status of these two perspectives.
Schema: This generalized, well-integrated knowledge about a situation, an event, or a person is called a schema. Schemas often influence the way we understand a situation or an event, and we can think of them as the basic building blocks for representing our thoughts about people Consistent: -Brewer and Treyens (1981): asked participants in their study to wait, one at a time, in the room pictured in this demonstration. Each time, the experimenter explained that this was his office, and he needed to check the laboratory to see if the previous participant had completed the experiment. After 35 seconds, the experimenter asked the participant to move to a nearby room. Then, the experimenter asked each participant to remember everything in the room in which he or she had waited. -The results showed that people were highly likely to recall objects consistent with the "office schema." Nearly everyone remembered the desk, the chair next to the desk, and the wall. However, only a few recalled the wine bottle and the picnic basket. These items were not consistent with the office schema. When time is very limited—in this case, 35 seconds in that room—people may not have time to process these schema-irrelevant items. Inconsistent: -However, we sometimes show better recall for material that violates our expectations. Specifically, people are more likely to recall schema-inconsistent material when that material is especially vivid and surprising. -Davidson (1994): asked participants to read a variety of stories that described well-known schemas such as "going to the movies." The results demonstrated that people were especially likely to recall schema-inconsistent events when those events interrupted the normal, expected story. -For example, one story described a woman named Sarah who was going to the movies. The participants were very likely to remember a schema-inconsistent sentence about a child who ran through the movie theater and smashed into Sarah. -In contrast, they were less likely to remember a schema-consistent sentence about an usher tearing the movie tickets in half and giving Sarah the stubs. -These results about schemas and memory may seem inconsistent. However, the outcome may depend on factors such as the details of the study and the length of the specific episode. In general, the results show the following trends: 1. If the information describes a minor event—and time is limited—people tend to remember information accurately when it is consistent with a schema (e.g., the desk and the chair in the "office"). 2. If the information describes a minor event—and time is limited—people do not remember information that is inconsistent with the schema (e.g., the wine bottle and the picnic basket). 3. People seldom create a completely false memory for a lengthy event that did not occur (e.g., the lecturer did not dance across the room). 4. When the information describes a major event that is inconsistent with the standard schema, people are likely to remember that event (e.g., the child who crashes into Sarah).
5. Discuss the basics of parallel distributed processing (PDP) models of semantic memory. In your answer you should: a. Distinguish between a local representation and a distributed representation, providing an example of each. Why are representations in PDP models sub-symbolic? b. Briefly describe how a PDP model works (just worry about the basics on this question). What does it mean that PDP models exhibit graceful degradation? c. Describe how knowledge is stored in a PDP model (see your text for additional information). d. Describe the concepts of default assignment and spontaneous generalization and how they are exhibited by PDP models. e. Based on your text and the lecture, describe two problems with PDP models.
a. -Local representation: in semantic networks, concepts are represented by nodes, which are symbols. This is local representation -Distributed representation: In PDP models, concept representation is distributed across nodes. It is a distributed representation -PDP models are intended to be similar to neuronal architecture -Ex: say you wanted to represent ice cream flavors. Chocolate would be local representation -A distributed representation would look like this (like a scantron) > The combination of nodes represents the concept > The nodes are sub-symbolic (they stand for symbols) > Meaning is in the pattern of activation b. -Input nodes are activated by some input --> input layer --> hidden layer --> output layer --> output nodes represent the concept -Ex: like machine learning (ex: eeg and cardiologists) PDP Models have been used to model: -Children's language learning -Decision making -The Stroop effect -Memory deficits in Alzheimer's Disease Strengths: -Can respond to new things (generalizes) -Mimics humans in many ways > a "lesion" does not totally shut the model down --> graceful degradation -The brain's ability to provide partial memory is called graceful degradation. 1. Cognitive processes are based on parallel operations, rather than serial operations. Many patterns of activation may be proceeding simultaneously. One process does not need to be completely finished before another process begins its operations. 2. A network contains basic neuron-like units or nodes, which are connected together so that a specific node has many links to other nodes. (Notice that this concept is captured in the alternate name for the theory: connectionism). PDP theorists argue that most cognitive processes can be explained by the activation of these networks 3. The process of spreading information from one node to other nodes is called spreading activation. As the name "parallel distributed processing" also suggests, a concept is represented by the pattern of activity distributed throughout a set of nodes. This view is very different from the commonsense idea that all the information you know about a particular person or object is stored in one specific location in your brain. 4. Consistent with the concept of situated cognition, the current context often activates only certain components of a concept's meaning. If you stroll past the meat department at your grocery store, you won't necessarily connect those plastic-wrapped items with the animal that clucks, pecks for food, and lays eggs. c. Crucially, these networks come to represent knowledge through training. That is, they are designed to loosely mimic connectivity among populations of neurons in the brain. Based on the principle that neurons that fire together also tend to wire together, the network learns to strengthen connections among relevant units in order to represent concepts and categories. Thus, through general computational and design principles, the network learns—as a function of training—how to structure itself to represent structure in the input to which it is exposed (and thus, in this case, to learn about concepts and categories). d. -One advantage of the PDP model is that it allows us to explain how human memory can help us when some information is missing. Specifically, people can make a spontaneous generalization by using individual cases to draw inferences about general information -For example, suppose that your memory stores the information in Figure 8.3 and similar information about other college students. Suppose, also, that someone asks you whether engineering students tend to be politically conservative. PDP theory suggests that the clue engineering student would activate information about the engineering students you know, including information about their political orientation. You would reply that they do tend to be politically conservative, even though you did not directly store this statement in your memory. Spontaneous generalization can also help to explain stereotyping -PDP models also allow us to fill in missing information about a particular person or a particular object by making a best guess; we can make a default assignment based on information from other similar people or objects -Suppose, for example, that you meet Christina, who happens to be an engineering student. Someone asks you about Christina's political preferences, but you have never discussed politics with her. This question will activate information in the network about the political leanings of other engineers. Based on a default assignment, you would reply that she is probably conservative. -Incidentally, students sometimes confuse the terms spontaneous generalization and default assignment. Remember that spontaneous generalization means that we draw a conclusion about a general category (e.g., the category "engineering students"). In contrast, default assignment means that we draw a conclusion about a specific member of a category (e.g., a particular engineering student). e. Problems: -Does it help us understand what humans do? -What parts of the model are important (like humans), what parts are irrelevant? -Ex: passport control task -Theorists emphasize that the PDP approach is most appropriate for tasks in which several processes operate simultaneously, and in which networks—just like brains—must learn about patterned input in one's environment that is highly structured. Indeed, pattern recognition, language, music, face recognition, and categorization, all involve input with a great deal of structure (such as, in English, the word "the" precedes nouns).
2. Describe the basics of semantic networks. In your answer, you should: a. Discuss the basic assumption of symbols in the semantic network. b. Describe what is meant by the terms node, link, and network. Next, specify how they combine form the basic architecture of semantic networks and discuss why these networks are referred to as semantic networks. c. Discuss the concept of activation and spreading activation in semantic network models. d. Take the example given in class of seeing a robin. Discuss, step-by-step, how semantic networks would explain our access to the concept robin and concepts related to it.
a. symbol: something that stands for something else -ex: moon stands for that thing in the sky Semantic networks assume that info is represented by symbols in our mind b. -Concepts are represented by nodes -Nodes are connected by links -All nodes and links are in a network -Nodes in these networks are sematically related making this a semantic network c. -All nodes have some level of activation > The mental activity of thinking of a concept -Nodes become active through stimulation from the environment or by other thoughts -Say that you saw a robin --> your robin node would be activated -This activation would spread to other, related nodes (spreading activation: spreads to related concepts) -The activation would keep spreading (ex: like a paper towel and water) -Semantic networks: think in groups, not isolation > we don't just access knowledge in isolation, we access other info too Evidence: -Accounts for typicality effects well > Ss faster to confirm that "A robin is a bird" than "A chicken is a bird" > Links for typical items (eg: robin) are stored more closely to a concept (eg: bird) than atypical items (eg: chicken) -lots of evidence from priming studies. d. We would think about robins and when that happens we would get access to other info too (eg: bird, red breast, blue eggs,...) in order (meaning that closer in map = less time)
9. Distinguish between the constructive model of memory and the pragmatic view of memory, describing the work of a.) Bransford and Franks (1971) and b.) Murphy and Shapiro (1994). What is the current status of these two approaches?
a. Bransford and Franks (1971): asked participants to listen to sentences from several different stories. -Participants were then given a recognition test that also included some new items, many of which were combinations of the earlier sentences. -People were convinced that they had seen these new items before. This kind of error is called a false alarm. In memory research, a false alarm occurs when people "remember" an item that was not originally presented. -Bransford and Franks found that people were especially likely to make false alarms when a complex sentence was consistent with the original schema. -For instance, they frequently made false alarms for sentences such as, "The tall tree in the front yard shaded the man who was smoking his pipe." -In contrast, participants seldom made false alarms for sentences that violated the meaning of the earlier sentences. -For example, they seldom said that they had heard the sentence, "The scared cat that broke the window on the porch climbed the tree." -Bransford and Franks (1971) proposed a constructive model of memory for prose material. -According to the constructive model of memory, people integrate information from individual sentences in order to construct larger ideas. -Later, they believe that they have already seen those complex sentences because they have combined the various facts in memory. Once sentences are fused in memory, we cannot untangle them into their original components and recall those components verbatim. -Notice that the constructive view of memory emphasizes that our cognitive processes are generally accurate. Sentences do not passively enter memory, where each is stored separately. Instead, we combine the sentences into a coherent story, fitting the related pieces together. We typically store an abstract of the information, rather than a word-for-word representation. -Your memory is typically accurate. However, the errors in cognitive processing can often be traced to strategies that are generally useful. In real life, a useful heuristic is to fuse sentences together. However, this heuristic can lead us astray if we apply it inappropriately. -As it turns out, the participants in Bransford and Franks's (1971) study used a constructive memory strategy that is useful in real life. However, it is typically inappropriate in a study that tests verbatim memory. b. Murphy and Shapiro (1994) developed a different view of memory for sentences, which they call the pragmatic view of text memory. -The pragmatic view of memory proposes that people pay attention to the aspect of a message that is most relevant to their current goals. In other words: 1. People know that they usually need to accurately recall the gist of a sentence. 2. They also know that they usually do not need to remember the specific wording of the sentences. 3. However, in those cases where they do need to pay attention to the specific wording, then they know that their verbatim memory needs to be highly accurate. -Murphy and Shapiro (1994) speculated that people are particularly likely to pay attention to the exact wording of a sentence if the words are part of a criticism or an insult. -From the pragmatic viewpoint, the exact words do matter if someone is insulting you! -In this study, participants read letters that presumably had been written by a young woman named Samantha. One group read a letter, supposedly written to her cousin Paul. In the letter, the hypothetical Samantha chatted about her new infant in a bland fashion. The letter included a number of neutral sentences such as, "It never occurred to me that I would be a mother so young". -A second group read a letter that was supposedly written by Samantha to her boyfriend, Arthur. Ten of the sentences that had been neutral in the bland letter to cousin Paul now appeared in a sarcastic context, though the exact words were identical. -For example, the sentence, "It never occurred to me that I would be a mother so young" now referred to Arthur's infantile behavior. -Murphy and Shapiro then gave both groups a 14-item recognition test that included (a) five of the original sentences, (b) five paraphrased versions of those sentences with a slightly different form, such as, "I never thought I would be a mother at such a young age," and (c) four irrelevant sentences. -People rarely made the mistake of falsely "recognizing" the irrelevant sentences. -However, correct recognition was higher for the sentences from the sarcastic condition than for the sentences in the bland condition. -People made more false alarms for the paraphrases of the bland sentences than for the paraphrases of the sarcastic sentences. -More generally, the researchers found that people were much more accurate in their verbatim memory for the sarcastic version (43%) than for the bland version (17%). -Perhaps we are especially sensitive about emotionally threatening material, so we make an effort to recall the exact words of the sentences. Current: -In reality, the constructive approach and the pragmatic approach to memory abstraction are actually compatible. -In many cases, we do integrate information from individual sentences so that we can construct a schema, especially when we don't need to remember the exact words. -However, in some cases, we know that the specific words do matter, and so we pay close attention to the precise wording. -If you are rehearsing for a play, or you are quarreling with a friend, you will need to remember more than the overall gist of a verbal message. -This conclusion about remembering both general descriptions and specific information is similar to our previous conclusions about semantic memory. -As discussed earlier, your semantic memory can store both general prototypes and specific exemplar-based information.
3. Discuss the evidence for semantic networks, citing: a. typicality effects b. Myer and Schvaneveldt's (1971) study of priming, describing the methods and basic findings. Why do their findings support the assumptions behind semantic networks?
a. Evidence: -Accounts for typicality effects well > Ss faster to confirm that "A robin is a bird" than "A chicken is a bird" > Links for typical items (eg: robin) are stored more closely to a concept (eg: bird) than atypical items (eg: chicken) -lots of evidence from priming studies. b. Myer and Schvaneveldt (1971): Ss see two strings of letters, respond "yes" if both are words, "no" if they're not -Lexical decision task (ex: valt-butter ; desk-mitten) -Two conditions of interest > related words: nurse-doctor = faster (855ms) > unrelated words: bread-doctor = slower (940ms) How would semantic networks explain this? -Don't have to travel as far -Spreading activation
10. Describe work on scripts and memory. In your answer you should: a. Define the term script, providing an example. b. Is recall of a script related to whether a script is identified in advance? Discuss, describing the work of Trafimow and Wyer (1993).
a. Script: A script is a simple, well-structured sequence of events in a specified order; this script is associated with a highly familiar activity -A script is an abstraction, in other words, a prototype of a series of events that share an underlying similarity. -The terms schema and script are often used interchangeably. However, script is actually a narrower term, referring to a sequence of events that unfold in a specified order -Ex: Consider a typical script, describing the standard sequence of events that a customer might expect in a traditional restaurant. The "restaurant script" includes events such as sitting down, looking at the menu, eating the food, and paying the bill. -We could also have scripts for visiting a dentist's office, for a trip to the grocery store, and for the first day of class in a college course. In fact, much of our education consists of learning the scripts that we are expected to follow in our culture b. In general, research demonstrates that people recall a script significantly more accurately if the script has been clearly identified in advance. -For example, Trafimow and Wyer (1993) developed four different scripts, each describing a familiar sequence of actions: photocopying a piece of paper, cashing a check, making tea, and taking the subway. -The researchers also added some details that were irrelevant to the script, such as taking a piece of candy out of a pocket. -In some cases, the script-identifying event was presented first. -In other cases, the script identifying event was presented last. -After 5 minutes of reading all four descriptions, the participants were asked to recall the events from the four original descriptions. -When the script-identifying event had been presented first, participants recalled 23% of those events. -In contrast, they recalled only 10% when the script-identifying event had been presented last. -As you might expect, the events in a sequence are much more memorable if you understand— from the very beginning—that these events are all part of a standard script
12. Describe work on gender-based inferences, focusing on a.) research in cognitive neuroscience and b.) data from the Implicit Association Test (IAT).
a. Using neuroscience techniques to assess gender stereotypes: -Osterhout and his colleagues (1997) assessed gender stereotypes by using a neuroscience technique. The event-related potential (ERP) technique records tiny fluctuations in the brain's electrical activity, in response to a stimulus. -Previous researchers had tested people who were instructed to read sentences such as, "I like my coffee with cream and dog." -As you would expect, the ERPs quickly changed in response to the surprising word dog, relative to the onset of the highly expected word (cream). -To examine gender stereotypes, Osterhout and his colleagues (1997) presented some sentences that would be consistent with gender stereotypes, such as, "The nurse prepared herself for the operation." -These stereotype-consistent sentences did not elicit a change in the ERPs. -In contrast, the ERPs changed significantly for stereotype-inconsistent sentences such as, "The nurse prepared himself for the operation." -In reading the word nurse, people had made the gender-stereotyped inference that the nurse must be female. -Consequently, the unexpected, stereotype-inconsistent word himself produced changes in the ERPs. -In a related study, White and her colleagues (2009) found similar results, with greater ERP changes for stereotype-inconsistent pairs of words. b. Using the Implicit Association Test to assess gender stereotypes: -Brian Nosek, Mahzarin Banaji, and Anthony Greenwald (2002): examined the gender stereotypes that mathematics is associated with males and that the arts are associated with females. -Suppose that they had asked their research participants (college students at Yale University) an explicit question, such as: "Is math more strongly associated with males than with females?" -When research is conducted with college students like these, the students are most likely to answer "No." -After all, when students are asked an explicit question like this, they have time to be analytical and to recall that a "Yes" answer would not be socially appropriate. -Instead of an explicit measure, Nosek and his colleagues used the Implicit Association Test -An implicit task asks people to perform a task, but the participants do not know what the task is supposed to measure. -The Implicit Association Test (IAT) is based on the principle that people can mentally pair two related words together much more easily than they can pair two unrelated words. -Specifically, when participants worked on the IAT, they sat in front of a computer screen that presented a series of words. -On a typical trial—where the pairings were consistent with gender stereotypes—the participant would be told to press the key on the left if the word was related to math (e.g., calculus or numbers) and also if the word was related to males (e.g., uncle or son). -This same participant would be told to press the key on the right if the word was related to the arts (e.g., poetry or dance) and also if the word was related to females (e.g., aunt or daughter). -Throughout the study, participants were urged to respond as quickly as possible, so that they would not consciously consider their responses. -When completing the IAT, people with strong gender stereotypes would think that math and males fit in the same category, whereas the arts and females would fit in a different category. -Therefore, their responses should be quick for this first portion of the task. -Then the instructions shifted so that the pairings were now inconsistent with gender stereotypes. -Now, on a typical trial, the participant would press the left key if the word was related to math and also if the word was related to females. -The participant would press the right key if the word was related to the arts and also if the word was related to males. -People with strong gender stereotypes should have difficulty associating math-related terms with women and arts-related terms with men. -Their responses should therefore be much slower for this second part of the task. -Nosek and his colleagues (2002) found that the students responded significantly faster to the stereotype-consistent pairings (the first task), compared to the stereotype inconsistent pairings (the second task). -In other words, most of these Yale students believed that math and males seem to go together, whereas the arts and females seem to go together. -They also analyzed the scores for women who strongly considered themselves to be feminine and strongly associated math with being masculine. -These women specifically did not associate themselves with mathematics. -Sadly, this anti-math tendency was even found with some women who were math majors! -In other words, gender stereotypes are not innocent cognitive tendencies. Instead, these stereotypes can have the power to influence people's self-images and their sense of academic competence. -The results showed that the countries with the highest scores on the IAT measure of gender stereotyping were also likely to have the highest "male advantage" scores on both the math and science test (TIMS). -Specifically, the correlation was +.34 for math scores and +.39 for science scores. -In other words, the countries with the highest measures of gender stereotyping were also more likely to be the countries where males performed better than females in both math and science. -Incidentally, the United States and Canada had average scores on both the IAT and the male-advantage measure.
1. a. Describe what is meant by an addressing system of information storage, providing an example. b. Distinguish between addressing systems of information storage and content addressable systems. c. Why is it apparent that our memory does not use an addressing system? Explain.
a. addressing system: each entity has its own unique address (computers do this) -in an addressing system our errors would be random -ex: who played Michael Scott in The Office? > address 78342, error 88342 > you might say Austin, Texas b. content addressable: the content of the memory is the storage address -as if you walk into a library and yell "Who was the lead actor on The Office?" -The book jumps out and runs to you -can explain our errors (and close calls, Will Ferrell in The Office) -Try answering this question: Does Phil Dunphy have a small intestine? > The answer is not in memory, no "book" to jump out at you -But...other info in memory can help you answer the question -The big question: How is our knowledge organized? c. Because in an addressing system our errors would be random