Representation of Knowledge

The Representation of Knowledge How do we organize and re-organize our mental representations and what cognitive processes guide this internal action? Copyright © Allyn & Bacon 2005 Overview  Knowledge:  The storage, integration and organization of information in memory  Knowledge is organized information  Knowledge is part of a system or network of information  Knowledge is “digested” information Models that attempts to explain how different types of information are encoded and processed by our cognitive system  Focus is on semantic organization since that is the primary code for LTM   Knowledge Representations:  Semantic structures allow us to identify what types of things are stored in the mind and how a stored thing is related to other entities in the mind Copyright © Allyn & Bacon 2005 Ways knowledge can be organized/represented  All with semantic underpinning-- knowledge may be represented in terms of:      Relationships Lexical (word) representations Propositional relations Images Neurological components Jennifer played basketball with her friends. Copyright © Allyn & Bacon 2005 Organization of Knowledge Knowledge may be represented in terms of: Relationship Lexical Representation (characters in source coded are translated into symbolic words) Propositions (smallest meaningful unit of Information about an assertion) Image Neurological Components Action Jennifer = woman Played = action Basketball= noun, sport Action= played Agent of action= Jennifer Object = basketball Parts of visual and association cortex, perhaps motor cortex Attribute Jennifer is a woman, she is tall, she is athletic Jennifer is a woman. Jennifer is tall. Jennifer is athletic. Parts of the visual and associative cortex, perhaps part of the right parietal cortex for face Recognition Parts of the visual, associative, motor and sensory cortex Spatial Properties Jennifer holds the basketball Jennifer has a basketball in her hands. Class Membership Jennifer is part of the following categories: woman, basketball player, tall people Woman, a type of which Jennifer is. Basketball players, a type of which Jennifer is. Parts of the association cortex Copyright © Allyn & Bacon 2005 5 Models of Knowledge Representation      Cluster Model (Bower)  Concepts tend to be organized in clusters based on meaningful associations Set-Theoretical Model (Meyer)  Concepts are represented in memory as sets or collections of information Semantic Feature-Comparison Model (Smith & Rosch)  Concepts are represented in memory as a set of semantic (meaningbased) features defining features & characteristic features Semantic/ Propositional Networks (Collins, Quillian, Loftus, Anderson)  Knowledge exists in memory as independent units connected in a network tied by relationships Neural Network Model (Rumelhart & McClelland)  Knowledge is represented in an organization of parallel networks– knowledge is the connections between processing units Copyright © Allyn & Bacon 2005 Clustering Model  Based on the belief that LTM is organized into semantic (meaningbased) clusters through functional associations Focus is on organizational factors that effect conceptual hierarchies Minerals Metals Stones  Rare Common Alloys Precious Masonry Platinum Silver Gold Aluminum Copper Lead Iron Bronze Steel Brass Sapphire Emerald Diamond Ruby Limestone Granite Marble Copyright © Allyn & Bacon 2005 Set-Theoretical Model  Semantic concepts are represented by collections of information (or sets)   Differs from clustering concepts because concept may be represented in LTM not only as exemplars (typical/best representations) but by attributes of the concept Instances of a category  Example: birds = canaries, finches, robins  Attributes of a category  Example: bird = cheeps, flies, has feathers Copyright © Allyn & Bacon 2005 Set-Theoretical Model  In this model, memory consists of constellations of events, attributes and associations where retrieval involves verification (a search through 2 or more sets of information to find overlapping exemplars  A robin is a bird   Compare attributes of one set [bird] with the attributes of another set [robin] Degree of overlap forms basis for decision Logical relationships= •Universal affirmative •All S are P •All Robins are Birds •Particular affirmative •Some S are P •Some animals are birds Robin Physical Object Living Animate Feathered Red-breasted ------------------- Bird Physical Object Living Animate Feathered ------------------------- P S S P Copyright © Allyn & Bacon 2005 Semantic FeatureComparison Model  Shares set-theoretical structure but differs in the following assumption:  Meaning of word can be represented as a set of semantic features along a continuum of importance  Robin = has wings, is a biped, has a red breast, perches in trees, likes worms, is untamed, is a sign of spring, etc.  Defining Features= wings, legs, red breast (these are critical!)  Characteristic features= perches in trees, likes worms, is untamed, is a sign of spring Copyright © Allyn & Bacon 2005 Semantic FeatureComparison Model  Validation of a relationship (A robin is a bird)  Defining features — essential (bird = robin)  Robin have beaks, wings, feathers Butterflies have wings  Characteristic features — accidental (bird = butterfly)   A true statement has both defining and characteristic features first stage of validation of the statement involves both defining and characteristic features of the two lexical categories (robin and bird). If there is considerable overlap = sentence validated. If the is no overlap = judged to be invalid. If there is some overlap, a second-level search is activated in which specific comparisons are made between the lexical categories on the basis of shared defining features weapon Copyright © Allyn & Bacon 2005 Semantic FeatureComparison Model  Support for this model (Rosch, 1977)  Some items are more typical of a category (more defining) then others Example =typical weapons Knife Gun Cannon Club Fist Chain  Category = Birds   Robin considered good example (prototype) Chicken or ostrich not     I saw a bird fly south Birds eat worms There is a bird in the tree Replace bird with robin or chicken, Robin shown to make more sense I heard a bird chirping on my windowsill Copyright © Allyn & Bacon 2005 Semantic FeatureComparison Model  Issues:  Collins & Loftus (1975) criticize using defining features as if they were absolute measures   Robin is still bird even if it has no wings/legs, can’t fly, etc. No single feature ―makes‖ a robin  So how separate defining feature from characteristic features Both Set-Theoretical Model and Semantic Feature Comparison Model share and are important for: •Providing specific information about multiple dimensions of semantic memory •Using systematic categorization as a starting point for theory of semantic memory Copyright © Allyn & Bacon 2005 Network Models Early Model  Knowledge exists in memory as independent units connected in a network by relationships  Early network Model of Collins & Quillian  Each word is depicted as a configuration of other words in memory, the meaning of any word represented in relationship to other words Semantic memory consists of vast hierarchical networks of concepts which are composed of units and properties and are linked by associationistic bonds   Bird and canary are stored in memory in terms of the relationship between them  Canaries are birds which are animals Copyright © Allyn & Bacon 2005 Network Models Early Model  Early network Model of Collins & Quillian Animal Has skin Can move around Eats Breaths Has fins Can swim Has gills Bird Has wings Can fly Has feathers Fish A canary is a subset of bird, bird has the property of flying, feathers-- So canary can fly, has feathers (does not need to be double stored—this system of semantic memory is economical) Canary Is tall Can sing Is yellow Ostrich Salmon Can’t fly Can bite Is dangerous Is pink Is edible Swims upstream Strength of this model: makes explicit the means by which information is retrieved from semantic memory Shark Copyright © Allyn & Bacon 2005 Network Models Early Model  Early network Model of Collins & Quillian Has skin Can move around Eats Breaths Has fins Fish Can swim Has gills Animal Can a shark move around? Assumptions of this model: •To validate an assertion, must search whole •network Takes time Criticism of the model: •associationistic strength varies within the network •Some associations violate the economy of the system Salmon Is pink Is edible Swims upstream Shark Can bite Is dangerous Copyright © Allyn & Bacon 2005 Network Models Semantic Network Where is Dearborn? Do fish have eyes? What is the square root of 9? Why do some women wear high-heeled shoes? Is an apple a porcupine? How many can you answer? How fast were your answers What influenced your answers? What accounts for differences? Copyright © Allyn & Bacon 2005 Network Models Semantic Network  Answer: A Semantic Network    Complex association network specific memories are distributed in a conceptual space with related concepts that are indicated by the length of connecting lines Differs from early model:  Got rid of hierarchy  Got rid of cognitive economy  Allowed links to vary in length to account for typicality effects  Spreading activation    Activation is the arousal level of a node Spreads down links Used to extract information from network Word list Copyright © Allyn & Bacon 2005 Network Models Semantic Networks  Note: Sleep is a false memory memory is constructive! Spreading Activation Theory  retrieval of one item triggers the retrieval of other, connected items as well  concepts are stored in memory in terms of their relationships to one another, so you can link to other concepts via these networks   SleepblanketLinusPeanutsLonestar Grill How did we get from sleep to the name of a restaurant?  Sleep triggered blankets (like on your bed), which for some may trigger Linus, Charlie Brown’s friend who carries a blanket, which triggered the name of the comic strip, Peanuts, which triggered the name of a local restaurant where patrons throw their peanut shells on the floor. Copyright © Allyn & Bacon 2005 Network Models Semantic Networks Priming (slip of tongue) More distant associations can be activated Show you the color red, likely you will recognize the word ―RED‖ faster = repetition priming, Further, see color red, recognize its associate such as fire engine or even more distant (red to fire engine to truck OR red to Sunset to clouds to sky) = semantic priming. Sunsets Clouds Sky Copyright © Allyn & Bacon 2005 Network Models Semantic Networks  Neurological Support (Posner & Colleagues):  From PET scans-- can distinguish brain patterns of activation between repetition priming (showing color green as prime for word green) and semantic priming (show color green for target grass)   Suggests visual word form is automatic and independent of attention (ventral occipital lobe) Suggests semantic priming works closely with attentional factors (left-lateralized) Copyright © Allyn & Bacon 2005 Network Models Propositional Notation  Most common notation used to represent meaningful information in Cognitive Psychology= Propositions  Propositions:  The smallest unit of knowledge that can stand as a separate assertion– it is the meaning underlying a relationship Lincoln, who was president of the United States during a bitter war, freed the slaves. •Lincoln was president of the United States during a war. •The war was bitter. •Lincoln freed the slaves. Research (Anderson, 1972) show that we preserve the meaning of the propositions and their relation in the complex sentence, rather than information about specific wording. Copyright © Allyn & Bacon 2005 Network Models Propositional Notation  Propositions are not only about the smallest meaningful unit, but also how units relates to other information– complex ideas broken down into simple relationships  Two widely used propositional notations:  Kintsch (1998)    Puts the relation outside parentheses with the entities related being listed inside Simple sentence: The boy went home.  proposition notation: went(boy, home) More complex sentences contain more than one proposition: The old professor gave a boring lecture.     gave(professor, lecture students) old(professor) boring(lecture) note that "students" in the first proposition is an inference here; it is not explicitly stated in the sentence Copyright © Allyn & Bacon 2005 Network Models Propositional Notation  Two widely used propositional notations:  Anderson (1983)– uses a visual network with entities and relation labeled by the roles they fill (here is this idea in an embedded parentheses form):  The old professor gave a boring lecture.  ((relation: gave)  (agent: old professor)  (object: boring lecture)  (recipient: students)) Kintsch showed that the number of propositions is a good predictor of the time to read and understand a paragraph, even when the number of words varies: A lecture that was boring was given by a professor who was old.  The earlier sentence (The old professor gave a boring lecture.) was 7 words and 3 propositions while this one is 13 words but with the same underlying 3 propositions  similar reaction times Copyright © Allyn & Bacon 2005  Network Models Propositional Networks Lincoln, who was president of the United States during a bitter war, freed the slaves. President of (Lincoln, US, war) Bitter (war) Freed (Lincoln, slaves) Time Each proposition represented by ellipses, connected by labeled arrows to its relation and arguments. Propositions, relations and arguments called nodes, arrows called links United States Object War Subject Agent Lincoln Agent Object Relation President of Relation Relation Slaves Bitter Freed Copyright © Allyn & Bacon 2005 The Fan Effect  Provides support for propositional networks (Anderson, 1974)  The more propositions you learn about something, the slower you are to retrieve any one of them    The hippie touched the debutante. The hippie kicked the policeman. The hippie kissed the prostitute.  In a propositional network, the number of links fanning out of the hippie node would vary with the number of propositions.  the more propositions learned about the hippie, the more links we would have to search during remembering and so the slower (and more errorladen) the retrieval would be. Copyright © Allyn & Bacon 2005 Paradox of the Expert   (Smith, 1980) Fan Effect results troubling because the more we know about something, then the more trouble we have remembering it. Called the "paradox of the expert" because experts know a lot about a topic, but seem to be able to remember it all comparatively easily.     The lawyer The lawyer The lawyer The lawyer climbed the ladder. read the paper. addressed the crowd. broke the bottle.  Yields the Anderson results: the more propositions learned about the lawyer, the longer it takes to recognize one later. Copyright © Allyn & Bacon 2005 Paradox of the Expert  Smith showed he could make the Fan Effect go away in the right condition:  Had another condition in which he gave another proposition about the lawyer which was a "punchline" that tied the other propositions together. In this case it was: The lawyer launched the ship.     In this condition, Fan Effect went away and time to recognize one of the statements was the same regardless of the number of ship-launching propositions learned as if there were only one unit of knowledge when there were many. What this statement does is tie the ship-launching statements together into a larger knowledge structure called a script -- in this case the script for launching ships. Thus, there is a paradox of the expert because the experts have larger knowledge structures like scripts to integrate individual propositions. One implication here is that when learning something new, we need to quickly integrate the details into more general structures or we will quickly get bogged down in the Fan Effect (this is captured by the saying "Can’t see the forest for the trees"). Copyright © Allyn & Bacon 2005 Schema Copyright © Allyn & Bacon 2005 Schema  Schemas  Abstract mental frameworks that hold typical supersets, parts and other attributes of a concept  Abstract models of the external world based on past experience   Based on a slot representation of knowledge that goes beyond semantic networks (goes beyond properties of concepts to give typical details) HOUSE       Isa: Building (points to the superset like in a semantic network) Parts: Rooms Also have another schemas within Materials: wood, brick, stone stored within rooms would be Function: Human dwelling information that they have Shape: Rectangular, triangular windows and ceilings (so can Size: 100-10,000 square feet infer that houses have windows   Materials, shape, etc. = slots Wood, brick, rectangular = default values and ceilings Key way we organize knowledge, allows us to make inferences Copyright © Allyn & Bacon 2005 Schema What is your schema for a soap opera? My schema for a soap opera: A show during daytime television that focuses on a group of dysfunctional individuals involved in "on again"/"off again" and/or passive/aggressive relationships resulting in relational angst, and typically consists of overly dramatic music and/or acting. 1) Someone is in trouble with the law. 2) Someone is missing or lost/ missing or secret child 3) Some type of identity confusion—either the person doesn’t know who s/he is or there is an unknown twin/look alike. 4) A new love starting or rekindled (or who is sleeping together). 5) A relationship that is in trouble/jeopardy (there are a several subschemata here – e.g., infidelity, doesn’t feel appreciated, thought the other person was dead, is being seduced by someone else, etc…) Copyright © Allyn & Bacon 2005 Scripts  Scripts: an abstract mental framework that describes typical sequences of action in a particular context.  Form of schema but about an event  also based on slot representation for typical attributes  Allow us to use a mental framework to act in appropriate ways depending on our role when we must fill in apparent gaps within a given context (quick adaptation)  Ex. Restaurant script  Allows us to make inferences when hearing sequences of actions from a familiar scene  Ex. Doctor visit Copyright © Allyn & Bacon 2005 Anderson’s ACT-R Model   Associationistic Model: Adaptive Control of Thought Model of knowledge representation and information processing Activation Declarative Memory Production Memory Storage Match Execution Retrieval Working Memory Outside World Copyright © Allyn & Bacon 2005 Anderson’s ACT-R Model  Production Memory  Very close to procedural knowledge knowing how Notion that underlying human cognition is a set of conditionalaction pairs called production rules or condition-action rules simplest form = If X Then Y   Foundation of ACT-R is Production Systems   Learn Propositions:  isa(t-test, statistical-test)  isa(F-test, statistical-test)  We also read and/or hear property propositions about these compares(ttest, two-groups)  compares(F-test, two-or-more-groups)  has(t-test, numerical-scores) Copyright © Allyn & Bacon 2005  has(F-test, numerical-scores) Anderson’s ACT Model  Production Memory = production rules or condition-action rules  propositional information tells us what can be done (e.g., t-test, F-test, etc.) and information about what can be done (e.g., t-test uses scores) but not tell us under what conditions they should be done and how to do them. For that we need conditionaction rules (i.e., IF-THEN production rules) like:       IF goal is to compare two groups AND each group member has a score THEN use t-test IF goal is to compare more than two groups AND each group member has a score THEN use F-test Copyright © Allyn & Bacon 2005 Connectionism/ Parallel Distributed Processing Network  Also called Neural Network Model   uses physiological processes of brain as metaphor for understanding cognition A neurally inspired theory of mind that suggests mental processes take place through a large set of simple units connected in a parallel distributed network  Differs from Previous Models:  Not simply sequential (like information processing models)– multiple processing happens “in parallel” – at the same time Not like other network models (where information was static copy representing a concept, proposition or type of information)   The information (concept, proposition, etc.) itself is not stored, instead what is stored is the connection between units– which allows neuron-like patterns to be recreated knowledge is stored in the strength of connection between units  Knowledge is distributed throughout the system within the connections between units (in terms of strength).  Units excite or inhibit each other throughout the system (each unit takes on a activation value and communicates with other units) this pattern of activation is knowledge! Copyright © Allyn & Bacon 2005 Knowledge: Other models = activation of the nodes, PDP = activation of the connections Connectionism/ Parallel Distributed Network Copyright © Allyn & Bacon 2005 Connectionism/ Parallel Distributed Network We start from the simple features at the bottom (from the earlier visual areas of the brain that are reached after the retina) and they are connected with both positive and negative weights to the letters in the middle layer). Thus, if a horizontal bar feature (on the left) is recognized, then it excites the letters consistent with that (e.g., A and T) while inhibiting the ones inconsistent with it (e.g., N). Once the letters are activated they inhibit each other and inconsistent words while exciting consistent words (e.g., ABLE). Once the words get excited they excite other consistent letters and the letters the other features. Thus, in word and letter recognition the excitation (positive activation) and inhibition (negative activation) flow from bottom to top and back. One seemingly paradoxical result of this is the word superiority effect. This effect is that recognizing a letter is easier (faster and more accurate) if it is part of a word than if it is presented by itself (Reicher, Journal of Experimental Psychology, 1969). This neural network shows why that would happen: i.e., because once part of a word begins to get recognize then it feeds back down the network to help excite letters consistent with it. If the words wasn’t there (i.e., if a letter was presented by itself) then this extra recognition help wouldn’t exist. Copyright © Allyn & Bacon 2005