認知心理学11. Semantic Organization: 意味的組織化

認知心理学もくじ

• 認知心理学もくじ

Today

• How do we store meaning?

What is semantic memory?

• Memory of meaning
• Based on relationships between concepts
• Organized hierarchically
• Will organized presentation lead to better recall?
• hierarchical organization:
  • Influences the amount of information retrieved
  • And the time required to retrieve information.

Recall: random vs. hierarchical: ''Bower, Clark, Winzenz, & Lesgold, (1969)''

• Ss presented with different nested category lists.
  • 28 words in each
  • Studied each hierarchy for 1 min
• IV: Hierarchical organization or random presentation
• DV: number of words recalled in a free recall task.

In the hierarchical organization condition the items were presented with the conceptual hierarchies you just saw.
In the random condition Ss saw the same words but they were inserted randomly within each spatial tree.
For example, if the conceptual hierarchies include instruments, plants, minerals, and body parts.
Each spatial tree would include items from each category randomly placed.

• hierarchical organization

• random organization

Recall: random vs. hierarchical: ''Bower, Clark, Winzenz, & Lesgold, (1969)''

• Results:

• A hierarchy can be used as a effective retrieval plan for cuing recall.

Sentence Verification Task

• The amount of time it takes a person to respond provides clues about the organization of semantic memory
  • “is an ostrich a bird?”
  • “can an ostrich fly?”
  • “is an ostrich a building?”
• sentence verification task
  • allow us to test how semantic networks are organized

Semantic Network Model: ''Collins & Quillian, (1969)''

• semantic network
  • model of organization where related concepts are arranged in a network
• This represents the way concepts are organized in the mind
  • Nodes
  • Links
• Features true of all animals stored at the top level.
  • Animals eat
• Features true of specific, basic-level categories are stored in the middle level.
  • Birds have wings
• Features specific to a certain species are stored at the bottom level.
  • Canaries are yellow
• A semantic network suggests that semantic information is organized based on relationships.

Effectiveness of Semantic Networks: ''Holley & Dansereau (1984)''

• Ss study a 3000-word passage.
• 1/2 Ss received training on constructing semantic networks.
• Test:
  • Multiple choice
  • Short answer & essay
• Using semantic networks is a good method for organizing knowledge.

Sentence Verification

• As those first studies demonstrated both hierarchical and semantic organization increase, the amount of information subjects are able to recall, and several theories have been proposed to account for this.
• Two theories:
  1. hierarchical network model: uses hierarchical relations for semantic info.
  2. semantic feature-comparison model: match test items to features in the category.

Hierarchical Network Model: ''Collins and Quillian (1969)''

• But if semantic information is hierarchically organized, it should also influences the time it takes to retrieve information.
• Two primary assumptions:
  1. Moving from one level of the hierarchy to another takes time.
  2. Having to retrieve features contained at each level takes additional time

• First assumption is that it takes more time to move between levels.
  • For example responding to the question "is a salmon animal?" should take longer then answering "whether a salmon is a fish".
• The second assumption states that it will take even more time if we have to retrieve the features contained within each level
  • Subjects should respond faster to salmon is a animal than to a salmon breathes.

• Ss respond T or F to property and category questions.

Category Questions	Property Questions
A canary is a animal.	A canary eats.
A canary is a bird.	A canary can fly.
A canary is a canary.	A canary is yellow.

• category question
  • ask about a set of relations. Is one category a member of another
• property question
  • ask about features contained within each category level.

• Hear:
  • A canary can fly. or A canary can sing.
• Verify:
  • A canary is a bird.

Problems

• Instances in which RT’s are not influenced by the levels of the hierarchy.
  • A Chimpanzee is a primate.
  • A Chimpanzee is a animal.

• the model predicts that Ss should respond faster to primate than to animal.
• However, responding is slower for primate rather than animal.

• Hierarchical model doesn’t account for the typicality effect.
  • Canary is a bird.
  • Ostrich is a bird.

• According to the typicality effect, typical members of a category should be easier to classify then less typical members.
• Thus, we should and do respond faster to canary is a bird rather then a ostrich is a bird.
• However, the hierarchical model predicts that there will be no difference in response times between the to because they are same distance away from the next level.

Feature-Comparison Model: ''Smith, Shoben, & Rips (1974)''

• The feature-comparison model attempted to overcome the limitations of the hierarchical model by proposing an alternative approach
• Assumptions
  1. Semantic meaning can be represented in memory as feature lists
  2. Classifications are made by comparing features rather than examining links in a network

Feature-Comparison Model: ''Smith, Shoben, & Rips (1974)''

Canary	Bird
Physical object	Physical object
Living	Living
Animate	Animate
Feathered	Feathered
Flies
Yellow
Sings

Feature-Comparison Model

• defining feature:
  • Features the object must have to be a member of the category
  • Distinguishes between two types of features
  • その物体の種に共通すること

• characteristic feature:
  • Features usually present but not necessary for category membership
  • 鳥の中でカナリアを区別するための特徴

Canary	types of feature
Physical object	D
Living	D
Animate	D
Feathered	D
Flies	C
Yellow	C
Sings	C

• Two stages:
  1. Compare all the features between the two concepts to determine similarity.
     1. If initial comparison reveals high or low similarity, then a decision can be made immediately
  2. If similarity lies between the two extremes examine just the defining features
     1. Needed when the similarity between the category concept and the example decreases

• To determine if a canary is a bird, we would compare all the features of a canary with all the features of a bird.
  • For example we often make very quick decisions about classification because the two items are highly similar.
• When two items are very highly similar, we have no need to focus on the defining characteristics within the second stage.
• In other words, we should be quicker to classify typical examples robin, sparrow and blue jay as birds.
• Then, examples that are less typical of the category like chicken, goose, and duck.

• What about negative examples?
  • A bat is a bird.
  • A bat is a bottle of wine.

• We see that when we have a highly typical positive example, a canary is a bird, we can make an immediate judgment without having to rely on the defining features of the category.
• Feature-Comparison Model predicts the opposite for negative examples that share similar features
• Bats are not birds, but they have certain features that are highly similar to birds.
• And because they share many feature, it is difficult to make a quick rejection during the initial feature comparison.
• Therefore the probability increases that you will engage in the second processing stage and examine the similarity between the concepts defining characteristics.
• So the feature-comparison model accounts for why some false statements are evaluated quicker then others which the hierarchical model failed to do.
• A negative response also.
• however, because the items are highly dissimilar, we can make an immediate judgment about classification which we could in the first example.

• category-size effect:
  • members of smaller categories can be classified quicker then members of larger categories
• So determining that a canary is bird, will be faster than determine that a canary is an animal.
• However, this also true for the hierarchical model.
• Classifying a canary as a bird takes fewer inferences then having to classify it as a animal.
• In other words, the example should be more similar to smaller category then to the larger category.
• However, there cases when the category size effect is violated.
• The feature model is able to account for these violations because its predictions are based on similarity rather than category size.

Limitations

• Relies on similarity ratings based on features to make predictions.
  • Weak predictions
• Classifications require computation
  • Uses features to detect the degree of similarity
• Little support for the use of defining features

• we have very little evidence that feature comparison is really what people do.
  • Therefore, the predictions made with this model are weak.
  • On the other hand, the hierarchical network model doesn’t even make these predictions.
• Computation was emphasized in the last chapter on the categorization of novel pattern. And it was essential for that.
  • But once we have classified a concept, is it really necessary to continue to compute the degree of similarity?
  • It would seem to be much easier to use associations among concepts.
  • For example, once we have learned that a robin is a bird, is it really necessary to compute the degree of similarity between the two every time.
  • It would be easier just to directly use the information you had learned.
• The model proposes that we use defining features only in the second stage and not the first.
  • However, we can argue that all features are more or less defining.
  • Even characteristic features can be used to define, and only the more defining features are used in the 2nd stage.
  • But again, there is almost no empirical evidence that subjects actually do this.
  • And there is some evidence that the opposite is true.

Limitations: ''Keil & Batterman (1984)''

• This smelly mean old man with a gun in his pocket came to your house one day and took your color television set because you parents didn’t want it anymore and told him he could have it. Could he be a robber?
• This very friendly and cheerful woman came up to you and gave you a hug, but then she disconnected your toilet bowl and took it away without permission and never returned it. Could she be a robber?

Limitations

• The first passage uses characteristic features of a robber and not defining features.
• The second passage uses defining features of a robber but not the characteristic features.
• Which person do you think the children thought was the robber?
  • One interpretation as to why is that characteristic features are more salient.
    • Characteristic features like mean and gun are directly observable.
    • However, a defining feature like taking without permission are more conceptual.
    • However, we do see a developmental trend towards the use of defining features as children get older.

Spreading activation models: ''Collins and Loftus (1975)''

• feature-comparison model has some advantages over the hierarchical model, but some limitations as well.
• The spreading activation model:
  • generates assumptions as to how activation spreads throughout the model.
  • A theoretical construct proposing that activation spreads from a concept in a semantic network to activate related concepts (CP 212)
  • The spreading activation model kind of took the best of both, and brought them together.
  • In other words, it incorporates both the use of semantic networks and feature computation.

• The spreading activation model is similar to the semantic network models, because of its emphasis on concepts joined together by links.
• However, in the spreading activation model, the length of each link represents the degree of semantic relatedness between two concepts.
  • For example sunsets are more closely related to sunrises and clouds than to the color red.
• This model also accounts for the typicality effect.
  • Shorter links suggest that the object is a more typical example of the category, while longer links indicate that the object is not a typical member of the category.

• Assumptions:
  • Activation decays with time
    • Street to cloud. Takes longer.
  • Activation decays with distance
    • red to house
  • Activation from a node is proportional to the number of pathways (fan effect)

• fan effect
  • 概念に伴う連合の数が多いと，再認課題においてプローブへの反応時間が増大する現象である。

Testing spreading activation: ''Meyer & Schvaneveldt (1975)''

• lexical decision task

Word or Nonword?
BALF
BUTTER
BREAD-> BUTTER
RIVER -/->BUTTER

• The important part is that people respond faster to butter when they had first seen bread rather than river.
• Spreading activation accounts for this because activation of a word should activate related words.
  • Butter should be activated by bread but not by river.
• Thus, activation of one word makes it easier to identify related words result in faster response times for related words but not unrelated words.

Testing spreading activation: ''Ratcliff & Mckoon (1988)''

• spreading activation predicts that activation of information continue beyond a single node.
• Lion and tiger are associated.
• And tiger is associated with stripe but lions aren’t associated with stripes.
• According to the model activation should spread form lion to stripes , but stripe should be less activated then tiger because activation decreases with distance.

-LION ---->(TIGER) -> STRIPE
-LADDER -> (???)--> STRIPE

• And this is exactly what we find.
• Subjects are faster to respond to stripe when lion is presented first than when ladder is presented first.
• Furthermore, subjects respond faster to tiger after being primed lion then responding to stripe after being primed with lion
• Activation can spread to more distant nodes

Limitations

• Too flexible.
• Hard to test the model
• Predictions of what will not happen are difficult to generate

• Advantage to semantic networks is that they are flexible.
  • We can introduce new assumptions when the current ones fail to account for empirical findings.
• One seemingly advantage to the spreading activation model is feature matching allows for the verification of semantic relations.
• However, the downside is that it is hard to test this.
  • Flexibility is useful to point. If a model becomes so flexible that we can use it to account for all the data; then, it losses its predictive power.
  • And a models predictive power comes from it being able to also predict what will not happen.
• Thus, it is hard to test this model because the assumptions inherent to it make it difficult to predict when something will not happen.
• However, people still like this model.
  • It is fairly simple to understand the assumptions are clear.
  • We can also use variants of it study other issues cognitive psychologists are interested such as how we integrate information

So Far…..

• amodal
  • the knowledge used in these systems are abstracted from our sensory experiences.
  • Knowledge that is abstracted from sensory experiences (CP 233)
• modal
  • knowledge is represented as sensory experiences (CP 233)
• They are not representative of the perceptual experiences we actually encounter

Amodal model

• Neurons in feature maps produce sensory representation.
• Perceptual states are abstracted into non-representational format.
• Feature lists and semantic networks.

Perceptual Symbol Systems

• assumes that verification of semantic statements occurs through mental simulations.
• are directly stored in LTM.
• modal approach because it stores sensory experiences such as audition, vision, taste, smell, and touch.
• receive information from LM by reenacting or simulating perceptual experiences.

「鷹が飛んでいる」という文を見せたあとに、翼を開いている鷹の絵と翼を閉じている鷹の絵の双方に対してマッチングタスクを行ったところ、 前者の方がその意味する状況に即しているので反応時間は速くなるという例(Zwaan et al. , 2002)。

• Form of grounded or embodied cognition.
  • Memories are mental simulations of past events- of how we have perceived things and interacted with them.
• Can also be used to explain sentence verification tasks

Modal systems

• Simulation is the reenactment of perceptual, motor, and introspective states acquired during experience with the world, body, and mind.
• As an experience occurs (e.g., easing into a chair), the brain captures states across the modalities and integrates them with a multimodal representation stored in memory (e.g., how a chair looks and feels, the action of sitting, introspections of comfort and relaxation

Zwaan, Stanfield & Yaxley (2002)

• The ranger saw an eagle in the sky.
• The ranger saw an eagle in the tree.

Zwaan & Yaxley (2003)

• Presented Ss with words pairs placed above or below each other.

Attic	Lake	Swan	Candle	Taxi	Nose
Basement	Boat	Sink	Flame	Peach	Mouth

• Results: subjects faster when word alignment matched alignment of our images.

• Their first study showed that we are faster to respond when the picture matches the sentence.
• This suggests that we have different representations for what objects should look like from different perspectives.
• They argues that is people do use images to help them make the decision; then, using the position of words should work as well.

Pecher, Zelenberg, & Barsalou

• Simulating:
  • is the reenactment of perceptual, motor, and introspective states acquired during experience with the world, body, and mind.
• As an experience occurs, the brain captures states across the modalities and integrates them with a multimodal representation stored in memory

• This suggests that subjects are simulating the perceptual information.
• If they weren't, subjects should not respond slower to shifts in modality.
• Semantic networks don’t predict this.
• Regardless of modality, it’s depend on the how far activation needs to spread to activate that node.

Glenberg & Kaschak (2002):

• Q: Does the sentence make sense?
  • もし意味が通るようならジョイスティックを上へ、そうでないなら下へ
• You closed the drawer.
• You opened the drawer.

Wu (1995)

• Ss list features of nouns (watermelon) or noun phrases (half a watermelon).
• Perceptual symbols (half a watermelon would reveal hidden features, red and seeds) predicts more features will be listed for the noun phrase
• Amodal systems predict no difference because half a watermelon doesn’t contain anymore features than a whole watermelon.
• Results
  • Ss listed many more characteristics of noun phrases over nouns alone.

Limitations of Perceptual Symbol Systems

• Not clear as to how information is integrated.
• Needs to combine perceptual information with the organization of semantic networks.
  • Organizing large bodies of knowledge is a strength of amodal theories such as semantic networks.

Clusters of Knowledge

• semantic network
  • show association but not how an object is used.
  • A theory proposing that semantic information is organized in long-term memory by linking concepts to related concepts (CP 212)
• perceptual symbol system
  • describe how objects are used but does not represent associations.
• Although semantic networks are good at organizing large bodies of knowledge, they do not represent clusters of knowledge so well

The ACT Model: ''Anderson (1976)''

• Assumptions:
  • Knowledge is stored in a semantic network.
  • Network consists of interconnected nodes.
  • Activation spreads from active nodes to activate new nodes and pathways
• Predicts that the more alternative paths activation can take the slower retrieval will be.
• Designed to be applicable to variety of cognitive tasks (scanning, inferences, retrieval etc..)
• Act assumptions a pretty much the same as assumptions in the semantic network.

ACT Model and Retrieval: ''Anderson (1976)''

• Ss studied sentences
  1. A hippie is in the park.
  2. A hippie is in the church.
  3. A hippie is in the bank.
  4. A captain is in the park.
  5. A captain is in the church.
  6. A debutante is in the bank.
  7. A fireman is in the park.

Anderson 1976

• IV- is the number of times a particular individual or location is mentioned.
  • Hippie = 3, Captain = 2, park = 3, bank = 2
• DV- is their reaction to test sentences
  • True: A hippie is in the park
  • False: A hippie is in the cave

• Did they see the sentence previously or not.
• RT increase as a function of the number of links to each person node or location node in the network
• So when they see the sentence, a hippie is in the park, it activates both hippie and park.
• Ss then must find the path within the network that joins both hippie and park together.

Problems

• The examples used by Anderson are not easy to integrate
  • There is no context, and thus no story
  • Relations are arbitrary
  • Furthermore, ACT suggests that the more knowledge you have, the slower retrieval will be.
• Maybe this is why retrieval is so slow and depends on the number links joining the concepts.

Schema Theory

• schema
  • framework for organizing clusters of knowledge.
  • A general knowledge structure that provides a framework for organizing clusters of knowledge (CP 228)
• Used to comprehend and store experiences.

• semantic networks are very good at demonstrating the way in which we organize knowledge together.
• However, the limitations of semantic networks is that they do not account for clusters knowledge.
• Schema theory is not a new idea.

Bartlett’s schema theory : ''Remembering (1932)''

• During this period the S-R approach to psychology dominated and not many people paid any attention to his theory
• With the reemergence of cognitive psychology in the late 50 and early 60’s, there was renewed interest in Bartlett’s theory.
• Even though it was born out of the behavioral period of psychology, there are fundamental differences between the S-R approach and Bartlett’s schema theory.

S-R approach	Schema Theory
Atomisitc	Molar
Associationistic	Nonassociationistic
Particularistic	Generic
Passive	Active

• Atomisitc vs. molar:
  • S-R approach is atomisitc and based on small units of knowledge (a single stimulus)
  • Schema theory is based on larger units of knowledge combined to form clusters.
• Assocationistic vs. non-associationistc.
  • S-R approach requires learning an association between the stimulus and response.
  • A schema on the other hand provides a structure for interpreting and encoding aspects of a particular experience.
• Particularistic vs. generic.
  • S-R approach looks a specific stimulus an the specific behavior it elcits.
  • A schema is more general and may contain an represent s a variety of particular instances.
• Passive vs active:
  • Learning an association between a stimulus and response such as in classical conditioning is very passive. It just happens
  • However, retrieving a schema from memory is a more active process because one must match a specific experience with the schema that best fits that experience

• Unfortunately, by 1970 Bartlett’s own students thought his theory was a complete failure.
• But in 1975, some American cognitive psychologists began arguing that schemas are in fact needed and not just in psychology.
• They were used to organize knowledge in artificial intelligence, linguistics, and motor performance.

Modern Schema Theory: ''Rumelhart (1980)''

• How knowledge is represented and how that representation facilitates the use of knowledge
• they took Bartlett’s theory an refined it.
• They kept most of the major assumptions, but were more specific as to what these knowledge structures looked like.
• And one way a schema facilitates the use of knowledge is through scripts.
• script
  • knowledge about how routine activities are accomplished.
  • Knowledge about what occurs during routine activities (CP 230)

• There are two possible ways in which a script is organized.

1. centrality of components

• • you think of the more important components before less important components.

• • For example, lets say you had to change to tire. Using centrality of components, we would expect our script to look like this.
  • This should seem very similar to the semantic network. However in this case, the distance of the links refers to how central that activity is to accomplishing your goal rather than a strong association between an object and its category.

2. sequential order

• • using the temporal or sequential order information in scripts are based on what comes first.

• Centrality vs. Sequential
  • If people use centrality, then they should be fastest verifying important activities.
  • If people use sequential order, they should be fastest verifying earlier activities.
• So we can generate predictions to test whether we use the centrality of components or sequential order when organizing scripts.

''Galambos & Rips (1982)''

• Ss rank order activities according to temporal or centrality
• Another group of Ss were given the name of the script (flat tire) and an activity (take out jack).
• Task: Time to decide whether the activity was part of the script.

• They found that Ss responded significantly faster to activities that were central to task, but there was no difference in speed of responses for earlier vs. later events.
• Supports the centrality of components organization of scripts.

Barsalou & Sewell (1985)

• Ss had 20 sec to name activities in a script.
  • Central condition: generate actions from the most central to the least central.
  • Sequential condition: generate actions from first to last.

• Results:
  • Central condition: 6.10 actions
  • Sequential condition: 8.17

• Barsalou and Sewell didn’t agree with these findings.
  • They believed the results were due to the nature of the task.
  • And that the results would be different with a different experimental procedure.
  • So subjects were able to generate more actions in 20 sec when recalling them in temporal order rather then based on centrality.
• What this suggests is that both centrality and temporal order influence our memory.
  • Central events provide quick access to the script
  • And temporal order is useful for retrieving the actions contained within a script.

認知心理学もくじ

• 認知心理学もくじ