lecture 8 - semantics Flashcards
-semantic
-semantic memory / representations
-lexicon
Semantic - to do with meaning (access/storage )
● Semantic memory / representations - a memory system for long-
term storage of facts
● Lexicon - our mental dictionary
lexical access
lexicon
lexical access : process of mapping sound onto meaning and retrieving information related to a given word: meaning but also grammar
lexicon :mental dictionary
models of lexical acess
cohort model
-explains how lexical acess happens for any word,
-model filters down the words it could be when hearing start of word etc
3 stages
connectionist model
-any of a class of theories hypothesizing that knowledge is encoded by the connections among representations stored in the brain rather than in the representations themselves
(there are 10 years between the models)
what is psycholinguistics
a discipline that describes psychological processes that
enable humans to learn and use (perceive and produce) language
Cohort and Connectionist models are classic psycholinguistic models
neurosicence of language
is a branch of cognitive neuroscience that
explores the neural mechanisms involved in language production,
perception, and acquisition
neuroscience of language can test psycholinguistic models…..
Neuroscience of Language can test Psycholinguistic models can introduce
new models that bind psychological and neural processes.
early evidence from semantic networks in the brain
sophie scot
-played people 4 different sounds
-clear speech, coded speech, (slightly degraded), reversed speech, and reversed and coded speech (both meaning and sound are degraded)
-looked at which part of brain lightsup when hearing these
-gyrus activated
-responds to sounds that have preserved the structure (apart from the revresible coded speech)
-if you go down further to superior temporal gyrus , lights up in the two conditions in which you understand what is being said
Early evidence from semantic
networks in the brain
binder et al 2000
-simialr to sophie scott experiment
-played 3 things : noise, tone, words
-can show where access to meaning happens (look at the different pcis)
neural evidence of semantic processing in the brain
when matched on sound complexity, words with meaning activate larger parts of temporal cortex → Lexicon and semantic memory system -
rooted in Temporal Lobes
models of semantic processing : the 2 requirements
(a) they need to express similarity - some words are related more than others
(cat→dog>cat→camel)
(b)need to support inference - we infer properties of concepts even when they are not
explicitly mentioned (if you see/hear/read ‘parrot’ you know/assume it can fly’).
major classes of models (models of semantic processing)
● Category-based theories
● Features-based theories
● Semantic space/network theories
semantic memory conatins many _____ and ________ categories
Semantic memory contains many discrete and independent categories
category theories
Category - basic-level natural language concept ‘cat’ or ‘tree’. Also ‘schema’/’prototype’
eg ‘seagul’ mapped to the bird category
-simplest model of semantic knowledge
● Comprehension - (1) finding the right category/ies for the input (2) retrieving specific semantic knowledge
● Conceptually similar items activate the same category representation
what evidence is there category theories
● Categories and words that refer to ‘prototypical’ members of the category are recognised, processed and learned faster by children – ‘apples’ are very close to ‘prototype fruit’
● We are more likely to substitute category members for category words when retrieving memories. ‘What he ate for breakfast?’ ’Some fruit’
● Words for ‘typical objects’ are learnt earlier than atypical ones
● Category specific semantic deficits (eg example of someone after a stroke forgot all musical instruments (that specific category)
category theory issues
● What about abstract concepts? Do they have categories/prototypes?
eg love / hate
● Categories diverge in context
-what you mean by ‘bird’ category might not activate the same one, eg penguins in antarctica, seagulls in scotland , context dependent
feature theories
Meaning is compositional -meaning for a word is stored as a collection of features (instead of accessing categories)
● A word activates associated features - properties that are likely to be true of that
item. ‘Bird’ - ‘can fly’
● Words/items are similar because of they
share features
● No need to access category information, no
explicit hierarchy of features
● Similar to ‘Schemas’
feature theory evidence
● Priming studies - words that share more features prime each other more
● Words that share features are more likely to be confused with each other (eg confuse cat and dog) (E.g. evidence aphasia patients - word substitutions such as he wants to says ipad but says a differnt appliance- so some features are shared but not in same category)
● Features and Category theories are not mutually exclusive (cant happen at same time)
feature theory issues
Many words don’t have obvious features they can be broken into. like abstract words E/g?
● Probabilistic ‘fuzzy’ features were proposed - but do they solve the issue? (like birds can fly but metaphorically can say other things fly)
latent space / network theories
● Conceptually similar to feature-based theories → meaning can be decomposed
● Instead of grounded features, there are latent dimensions and each word activates across multiple dimensions and has some probabilistic activation value (more/less)
this theory allows learning (contrast to two other theories)
● Word meaning - a vector though all latent dimensions.
● Latent dimension - black box, similar to neural network layers that encode arbitrary characteristics of the items/words that are
critical for a given task
network theory - evidence
Can be very powerful - since they are mathematically specific, while feature and categories are arbitrary
● Latent dimensions NNs outperform features-based NNs in different tasks
● Activation form layers from speech recognition NNs (like GPT2) - more in
last lecture
network theory - issues
In experimental ‘human’ evidence hard to distinguish between features and latent dimensions.
● There is neuroimaging evidence for grounded approach - that dimensions
of meaning are not arbitrary - embodied cognition (more below)
evidence for priming
Hutchinson et al., 2003
-can category or feature models explain ‘mediated’ priming
mediated priming : Mediated priming refers to the activation of a target (e.g., stripes) by a prime (e.g., lion) that is related indirectly via a connecting mediator (e.g., tiger).
Neither category or feature models can explain the case of ‘mediated priming’
Mediated priming – ‘LION’ primes ‘STRIPES’, even if Lions do not have stripes
When ‘LION’ is activated, activation spreads to include ‘TIGER’ and tigers have stipes
Network theories – partial explanation
Neural evidence for semantic categories and features
Luizziet al., 2020
fMRI brain response to written
words. Each word - activity patten
● Words from the same category show similar activity → large network frontal, temporal and parietal lobes
● Words that share the same
specific features (size, motion)
are activated similarly → more
localised - temporal areas
embodied semantics
-embodiement
Embodiment: semantic understanding is grounded in bodily experiences
embodied semantics
-simulation
Simulation: understanding concepts/words involves mental simulation of sensory
and motor experiences associated with those concepts.
embodiment semantics ‘-context dependency
- Meaning is often understood relative to the situation or
context in which it occurs.
embodiement semantics
-action and perception integration
a key feature is integration of action and perception systems in semantic processing, where understanding language involves activating relevant sensory-motor representations
smeantic representations are embodied
Tomacello et al., 2017
Words - a distributed collection of memories (auditory, motor, visual)
No ‘one place’ - activity pattern in the network
Multimodal hubs - joining information from different modalities
