Representing concepts Flashcards
What do all models of language perception assume about word meaning?
All models of language perception assume that meaning is stored/ represented in semantic memory
>Our “mental lexicon” connects abstract words with their underlying meaning
Describe the results of a study which examined evidence for an existence of such a mental lexicon. What does this suggest?
Examined whether the N400 for semantic word pairs (traffic-jam), associative word pairs (rabbit- sheep), both and neither. Semantic relation produced a same larger N400 as unrelated words while association word-pairs drew a smaller N400 the same as them both combined.
This suggests that we built these concepts of semantic networks where meaning is represented in nodes within an interconnected network. These models usually rely on pre-programmed associations, they don’t learn themselves.
Describe a type of model in which these models can be contrasted against
High dimensional semantic space accounts such as models like HAL (hyperspace analogy to language) and LSA (latent semantic analysis). In these models the meaning of each individual word is represented in how often that word occurs in conjunction with other words. This is derived from language corpora (newspapers, blogs etc). E.g hammer often co-occurs with nail, board, mallet, not likely to co-occur with chromosome. In this fashion these models can learn themselves and adapt. There are matrices with word associations and each word has its own vector.
Describe an early study which lends support to the effectiveness of these high-dimensional semantic space accounts models and the original results
An old priming experiment in which participants were presented with little stories in which a homophone was utilised in two different ways (e.g gardener getting rid of a mole with water, doctor getting rid of a mole with surgery). They were interested to see whether this would prime the either the association: ground (animal) or face (spot), or whether is would activate the inference: drown (animal) or cancer (spot).
The results showed that when the word mole was primed, both meanings for the homophone was acrtivated (ground, face) immediately after the prime but one second after the prime, a gap grows between them with the appropriate prime producing a faster response. This trend was also the case for the inference words.
How does this early priming study lend evidence to these HDSSA models?
If you apply models such as LSA which have been trained to these stories/ snippets used in the studies and then evaluate the vectors which come with either the semantic target that fits or doesn’t fit the contextual meaning of the word mold and with the inference which fits or doesn’t the context then you see a similar result to the human reaction times. The strength of the relationship between the vector representing ground and the vector representing the relevant context is larger than for the association between the vector representing face and the vector for that same context for example.
This suggests that a well trained LSA model can show output similar to how humans behave when they read sentences like this and could be an indication of how we learn and store meaning.
Neurologically, what assumptions do both semantic network models and HDSSA make?
Meaning is (probably) stored as patterns of inhibition and excitation and this depends on the strength of connections between neurons.
These neurological assumptions could suggest that meaning in simply a set of rules. What philospohical problem is related to this?
The chinese room experiment explaining why AI cannot be considered as intelligent simply by following rules
There are parallels which can be drawn between this problem for AI not being intelligent and how this proposes we learn language:
Written or spoken language is perceived
>Representation of physical stimulus
This activates certain neurons in the brain
>Representation of meaning
These neurons then in turn activate other neurons
>Response
What do thought experiments such as these attempt to do ?
Not meant to be executed
Thought experiments test concepts
>Search for consistency and contradictions
>Make use of intuition and logical reasoning to test hypotheses
How has this problem been adapted to a problem in cognitive neuroscience?
=> Spoken / written words are like Chinese characters entering the Chinese Room
=> Meaning representation in the brain is like John Searle in the Chinese Room
=> Connections between cells in the brain function as the rule book in the Chinese Room
According to the analogy of the chinese room, meaning will not actually arise within this system.
This is known as the (symbol) grounding problem in language
Describe two similar solutions to the grounding problem
Grounded/ embodied cognition: meaning is not symbolic/ abstract but anchored in experience
Indexical hypothesis: Mental representation is not abstract but directly connected (indexed) to the physical world.
Describe a study which directly compares the indexical hypothesis to HAL/ LSA models
If we we believe that HAL and LSA are a good representation of the neural components of assigning meaning to words the words like jumper are not related to words like leaves or water as they rarely co-exist together.
the indexical hypothesis on the other hand, says that our experience with the world makes a difference. So that if you read the sentence “Marissa forgot to bring her pillow on her camping trip. As a substitute for her pillow, she filled up an old sweater with leaves” then our experience with the world would allow this sentence to have a quicker read time than if the word was substituted with the word ‘water’.
This is the case rather than them having similar reading times as would be predicted by the LSA/ HAL, therefore it follows indexical hypothesis.
What kind of predictions would grounded cognition make about word meaning in our brains?
Since meaning is not abstract and directly tied to our experiences, there should be modality specific experiences for a given concept e.g a banana. The concept of banana in our brains would be tied to the colour of a banana, the shape, smell or taste of a banana and how a banana is handled or moved around.
Describe a study which explores the modality of colour in this grounded cognition hypothesis
We know that colour activates brain area V4 in the occipital lobe. Simmons et al compared a perceptual task in which chromatic (coloured) / achromatic stimuli with a conceptual knowledge task in which participants were asked whether the underlying concept of specific words had that specific colour in them.
When brain scans were carried out while participants were carrying out tasks, it was shown that there was seperate brain activation for each task while also overlapping considerably. This suggests that a word which just has tasks with no physical colours shown does activate colour representations in the brain at a very perceptual/ basic level.
Describe a study which explores the modality of shape in this grounded cognition hypothesis
We know from wider (low level visual) research that shape is processed in the fusiform cortex. Wheatley et al used words from 20 different categories (e.g birds, canines). These were repeated sequentially and the task was to judge the relation between the words; whether they were identical, related but different (within a category; dog, lion) or unrelated (between category; mango, chair).
A key similarity between related words for this study is that they shared a similar shape. Repetition suppression was utilized to determine whether the second word produced a smaller activation in the brain region. The suppression was largest in the identical situation but also present in the related situation, indicating that the shape concepts is activated by the word and this is what leads to repetition suppression in the fusiform gyrus.
Describe a study which explores the modality of smell in this grounded cognition hypothesis
Smell is processed in the bilateral orbitofrontal cortex. Goldberg et al used words belonging to certain categories (fruits, birds, body, clothing and control) as stimuli. They found that fruit-words but not other category words activates the orbitofrontal cortex.
