Theme 3: Semantic Memory Flashcards
What type of amnesia relates to impaired episodic memory?
hippocampal amnesia
what form of impairment is caused by the degeneration of the anterior temporal lobes?
semantic dementia
what is semanticisation?
when EMs changing to SMs over time
describe the hierarchical network theory of concept organisation
SM is organised into a series of hierarchical networks. Property info is stored as high up the hierarchy as possible to minimise the amount of info needing to be stored in SM. Processing takes longer when concepts are many levels apart
What are the limitations of the hierarchical network theory of concept organisation?
- Familiarity explains processing speed (not the distance between the concepts)
- Typicality effect also counters theory
- It falsely assumes that concepts belong to rigidly defined categories (they dont bc of ambiguity & vagueness)
What does the spreading activation theory state?
SM is organised on the basis of semantic relatedness. A concept is activated when brought to our attention. Activation spreads to other concepts, where closely semantically related ones are more greatly activated
What evidence supports the spreading activation theory?
- the typicality effect
- semantic priming leads to greater concept activation
is performance the fastest and most accurate at the superordinate, basic, or subordinate level?
the superordinate level
how do we know that different kinds of information about an object are stored in different brain locations?
because of patients with category-specific deficits
describe the hub-and-spoke model
Spokes: the 6 modality-specific brain areas
Hub: a general, modality-independent unified conceptual representation where we integrate knowledge of any given concept
where are hubs located according to the hub-and-spoke model?
in the anterior temporal lobes
Describe the Controlled Semantic Cognition framework proposed by Ralph et al. (2017)
Semantic cognition (SC) relies on two principal interacting neural systems:
Representation → encodes knowledge of concepts through the learning of higher-order relationships
Control → manipulates representational system activation to generate inferences and behaviours that are appropriate for each specific context
in which ways do we see that the ATL’s function is graded (Ralph et al., 2017)?
from its functionally graded cytoarchitecture, the graded partially overlapping connectivity across ATL subregions, & its graded functionality
what do we see happens moving away from the central point of the vlATL (Ralph et al., 2017)?
moving away from the central point of the vlATL weakens the cross-modal semantic function of the ATL and is more tied to a specific input modality
One could argue that the graded functionality reflects multiple mutually exclusive ATL subregions. Is this true or not, and why (Ralph et al., 2017)?
it is false because units anatomically closer to a given modality-specific spoke take part in all types of semantic processing, despite contributing somewhat more to tasks involving the proximal modality
Which area in the ATL is responsive to visual materials and concrete concepts (Ralph et al., 2017)?
the medial ATL
Which area in the ATL is responsive to social concepts (Ralph et al., 2017)?
the polar and dorsal ATL
Which area in the ATL is responsive to auditory/verbal stimuli and abstract concepts (Ralph et al., 2017)?
the anterior STS
research in various semantic impairments led to two important updates in the new connectivity-constrained version of the hub-and-spoke model. Which are these (Ralph et al., 2017)?
- semantic representations reflect collaborations between hub and spokes
- modality-specific info will be differentially important for some categories, which is analogously reflected in impairments to modality-specific regions
Semantic cognition control is implemented within a ____ neural network that interacts with (but is separate from) the ____ (Ralph et al., 2017)
distributed
semantic representation network
In their paper, Ralph et al. (2017) describe an integrative system for semantic control. What are its components and how does it work?
there are two aspects:
1. modality-specific info interacts with a transmodal hub to form generalisable concepts
2. as part of a separate executive control network, a region represents the current task context
drawing from these 2, the integrative system generates task-, time-, and context-relevant behavioural responses
What is semantic aphasia and how does it occur (Ralph et al., 2017)?
it is a deficit in manipulating and using semantic knowledge, and occurs from prefrontal or temporoparietal lesions
how does semantic aphasia prove the existence of a separate semantic control network, as described in the framework of Ralph et al. (2017)?
prefrontal regions do not encode semantic representations but are crucial for accessing, retrieving and executively manipulating semantic knowledge. Therefore, damage to those regions will cause deficits in semantic knowledge manipulation, rather than loss of semantic knowledge
Is functional specialisation graded in the control network (Ralph et al., 2017)?
there is superior–inferior functional specialisation:
inferior regions boost retrieval of weakly encoded information and superior regions contribute to more domain-general control
What did the hierarchical processing model by Quillian propose (McClelland & Rogers, 2003)?
it proposed that if concepts are organised into a hierarchy progressing from specific to general categories, then propositions true for all members of a superordinate category could be stored once at the level of the superordinate category
What is the perceptual-to-conceptual shift (McClelland & Rogers, 2003)?
A hypothesised developmental transition where infants initially categorise objects on the basis of their directly perceived visual properties, but later come to categorise them on the basis of deeper relationships
What is the premise of the parallel distributed processing (PDP) model (McClelland & Rogers, 2003)?
processing takes place via the propagation of activation among simple, neuron-like processing units
What is pattern completion in the PDP model (McClelland & Rogers, 2003)?
semantic information is reconstructed in response to probes
What is meant by McClelland and Rogers (2003) when they state that, according to their PDP model, items have internal representations?
When a network is presented with a given input, a pattern of activity arising across its hidden layer is the internal representation of that input
What aspect of semantic cognition does PDP accurately represent (McClelland & Rogers, 2003)?
its graded nature
Rumelhart built on the PDP model, emphasising its feedforward nature. How is this apparent in the model (McClelland & Rogers, 2003)?
activation flows from units representing items and relations, through hidden layers, to an output layer containing units corresponding to possible completions of three-constituent propositions
What is a limitation of Rumelhart’s feedforward PDP network proposition (McClelland & Rogers, 2003)?
it only learns through gradual learning, which doesn’t account for our ability to learn quickly. Furthermore, increasing the network’s learning rate leads to catastrophic interference
What is catastrophic interference (McClelland & Rogers, 2003)?
Rapid learning in ANNs causes large changes in connection weights that cause previously learned material to be destroyed
How does the Complementary Learning Systems Theory combat the caveats of the Feedforward PDP Network proposition (McClelland & Rogers, 2003)?
it comprises of a slow-learning semantic knowledge system which is complemented by a second, fast-learning system in the medial temporal lobes.
How does the PDP model resemble true human semantic learning (McClelland & Rogers, 2003)?
like children, the network differentiates concepts progressively, but names first at an intermediate level, overextending frequent names during an intermediate stage of development
What is coherent covariation (McClelland & Rogers, 2003)?
the consistent co-occurrence of a set of properties across different objects (it generally refers to the co-occurrence of more than two properties)
What is the importance of coherent covariation following the PDP model (McClelland & Rogers, 2003)?
because progressive differentiation arises from the coherent covariation of attributes across concepts, & the connection weights that determine the representation and processing of concepts tend to be driven in the same direction by their shared properties
How do various categorical variables arise in the PDP model (McClelland & Rogers, 2003)?
through multiple waves of differentiation
What causes the PDP network to produce intermediate level names before names at other levels (McClelland & Rogers, 2003)?
(a) differential exposure frequency and (b) the pattern of covariation of properties across concepts
Why might some concepts might be better or more coherent than others, and some properties might be more central to a concept than others, according to McClelland & Rogers (2003)?
Because coherent covariation of properties (a) leads to the overextension of properties to objects that do not have them, and (b) determines the strength with which a given feature contributes to representational change in a single learning episode. Properties that covary together generate larger weight changes throughout the network, and thus more strongly influence the development of internal representations
Why do people generalise properties differently, depending on the type of property and the type of concept to which it is applied, according to McClelland & Rogers (2003)?
Differential generalisation arises when there is domain-specific covariation of properties in experience
Children experience reorganisations of conceptual knowledge throughout their development. How is this modelled in the PDP model by McClelland & Rogers (2003)?
as it acquires information about other types of relations, the coherent covariation of the original information dominates learning, and the internal representations reorganise to capture underlying taxonomic organisation rather than the appearance information
What are feature listing tasks used for in psychology (De Deyne et al., 2017)?
to empirically measure how meaning is represented in small parts of the lexicon and predict measures like relatedness and typicality
What is the thesaurus model in linguistics (De Deyne et al., 2017)?
a method to study the mental lexicon, where the basic unit within WordNet is a synset - a set of synonymous words
what is latent semantic analysis (De Deyne et al., 2017)?
a corpus approach to studying the mental lexicon that captures word meaning by comparing how similar the contexts are in which two words occur
What is the word association task and what is the benefit of using it to study the mental lexicon (De Deyne et al., 2017)?
it is an empirical measurement tool for accessing semantic representations
the benefit is that it utilises free association, thus, it doesn’t limit people’s responses
How is the continued word association paradigm different to the word association task, and what are the benefits of using it (De Deyne et al., 2017)?
in this paradigm, people provide multiple associates to each cue
This approach is advantageous because (a) weaker associations can be collected and (b) the resulting network representations are denser and thus more suited to capture the distributional properties of meaning
How can spreading activation be represented in an ANN (De Deyne et al., 2017)?
as a stochastic random walk defined over the network
What is represented in the macroscopic, mesoscopic, and microscopic levels of a semantic network, respectively (De Deyne et al., 2017)?
- the entirety of the network
- a subset of nodes
- a single node & how it connects to the rest of the network
What does the semantic network structure look like at the macroscopic level (De Deyne et al., 2017)?
The mental lexicon networks contain a small number of highly connected nodes or hubs. These hubs exhibit a degree (i.e. number of connected nodes) that is much higher than other nodes, making the networks robust against damage and allowing for efficient information distribution
explain the mechanism of differential attachment (De Deyne et al., 2017)
it is an ANN growth principle that assumes new nodes become connected to the network proportionately to the number of existing connections they have with neighbouring nodes
explain how the mechanism of preferential atachment relates to real life human learning behaviour (De Deyne et al., 2017)
it explains how the age of word acquisition and its frequency in language independently contribute to the ease with which a word is processed
What is stated in the classical associative theory of creativity (De Deyne et al., 2017)?
creative individuals have a richer and more flexible associative network; thus, they may have more associative links in their network and can connect associative relations faster
what information can we derive from examining the mesoscopic level of the mental lexicon (De Deyne et al., 2017)?
information about word meaning and the processes and parts of the network that are involved in retrieving it
What observation about semantic networks is encompassed in the free categories organisation principle (De Deyne et al., 2017)?
network clusters consistently show a widespread thematic structure
What was described in the early propositional network model by Collins and Quillian (1969) (De Deyne et al., 2017)?
closeness between a pair of nodes in terms of the paths connecting them predicts RTs for verifying sentences
How do weak links in an ANN affect predictions for different kinds of words and semantic relations (De Deyne et al., 2017)?
introducing weak links represents a systematic improvement over networks derived from single-response procedures
How do indirect links in an ANN affect predictions for different kinds of words and semantic relations (De Deyne et al., 2017)?
they contribute to the relatedness of word pairs, improve predictions of human similarity judgments, and extract categorical relations between words
How does an ANN’s directionality affect predictions for different kinds of words and semantic relations (De Deyne et al., 2017)?
when undirected networks are derived, the density of the network increases. Increasing density improves prediction, but ignoring directionality hampers prediction of similarity judgments
Wht does associative priming reflect in an ANN (De Deyne et al., 2017)?
it reflects the presence of an associative link and the strength of the links between nodes
Research involving EM effects (e.g., false memories, recognition, cued recall) have been used to assess the underlying structure supporting information retrieval from the lexicon. What did this line of research find (De Deyne et al., 2017)?
that direct and indirect paths between words affect the intrusion of non-presented items in false memories and efficiency in cued recall and recognition tasks
What is node centrality and how does it inform us (De Deyne et al., 2017)?
it is the number different connections (set size) of a word, and it explains why certain words are processed more efficiently than others
what is the fan effect (De Deyne et al., 2017)?
the more things one learns about a word, the longer it takes to retrieve those facts
what is the concreteness effect (De Deyne et al., 2017)?
highly imageable words will be processed faster and more accurately than abstract ones
Which cognitive phenomena can be assessed at the macroscopic level of semantic networks (De Deyne et al., 2017)?
language development, creativity, and communication and thought disorders in clinical populations
Which cognitive phenomena can be assessed at the mesoscopic level of semantic networks (De Deyne et al., 2017)?
lexicon organisation principles, semantic relatedness, semantic priming and word retrieval processes
Which cognitive phenomena can be assessed at the microscopic level of semantic networks (De Deyne et al., 2017)?
word processing advantages
what is the aim of multimodal models of concrete concepts (De Deyne et al., 2021)?
they combine linguistic & sensory representations to determine to what degree linguistic representations capture sensory properties
Why do abstract concepts pose a problem for embodiment views of word meaning (De Deyne et al., 2021)?
because some abstract concepts do not have clearly identifiable referents. This contrasts embodiment views that state that both concrete and abstract concepts are grounded in the same systems that are engaged during perception, action, and emotion
How do distributional semantic views explain how abstract concepts are represented (De Deyne et al., 2021)?
abstract concepts are represented in terms of their distributional properties in language
What does the Affective Embodiment Account (AEA) posit (De Deyne et al., 2021)?
that abstract concepts are grounded in internal affective states. The grounding is quite broad, covering not just abstract words for emotions but also words that evoke an affective state
which 2 types of language models did De Deyne et al (2021) compare?
word association models and distributional linguistic models
How do word association models assess word meaning (De Deyne et al., 2021)?
word meaning is measured as a distribution of respectively weighted associative links encoded in a large semantic network
how do distributional linguistic models assess word meaning (De Deyne et al., 2021)?
word meaning derives from word co-occurrences from large text corpora. Semantic representation is captured from distributional properties in the language environment without integrating it with info from other modalities
In their study, De Deyne et al. (2021) compared how well each of the models predicted human similarity judgments. How did they use the triadic comparison to do that?
Participants rated similarity by picking the most related pair out of three words. There were two conditions, one in which the words were concrete and one where they were abstract.
Which model’s performance was significantly improved with the addition of visual information (De Deyne et al., 2021)?
the distributional linguistic model
How did adding affective information impact the performance of the two models in the study by De Deyne et al. (2021)?
it improved the performance of the distributed linguistic model. It didnt impact the word association model
Adding affective information improved model performance for ____ words, and adding ____ information improved performance for ____ words. This was only observed for the ____ model (De Deyne et al., 2021)
abstract
visual
concrete
distribution linguistic
What was the main finding of the study by De Deyne et al. (2021)?
the word association model of meaning captures visual information in concrete words and affective information in abstract words
consistent with AEA, the distributional linguistic models were most improved by adding affective information