The document discusses several classic approaches to understanding the meaning and representation of words, including semantic networks, semantic features, prototype and exemplar theories, and how concepts are combined. It also covers figurative language like metaphor and idioms, as well as connectionist approaches and the neuroscience of semantics. Key topics include how semantic memory is organized, how word meanings relate to concepts and categories, and experimental evidence from tasks like sentence verification.
2. Out line
Introduction
Classic approaches to semantics
Semantic Networks
Semantic Features
Family Resemblance Models
Combining Concepts
Figurative Language
The neuroscience of semantics
Connectionist approaches to semantics
2
3. Hello!
I am Ata Mohammed Saeed
I am here because I love to give presentations.
You can find me at atta.saeed@gmail.com
3
4. Introduction
How do we represent the meaning of words?
How do we organize our knowledge of the world?
Word meaning, examines issues involved in the study of semantics, in
particular how we represent the meanings of individual words.
Categorization, associations between words, use of metaphor and idiom,
and connectionist modeling of semantics are among the topics addressed.
5. First, we can translate words from one language to another, even though not every
word meaning is represented by a simple, single word in every language.
Second, there is an imperfect mapping between words and their meanings such that
some words have more than one meaning (ambiguity), while some words have the
same meaning as each other (synonymy).
Third, the meaning of words depends to some extent on the context. Hence a big ant
is very different in size from a big elephant, and the red in “the red sunset” is a
different color from “she blushed and turned red.”
5
(Hirsh-Pasek, Reeves, & Golinkoff, 1993)
6. Tulving (1972) distinguished between episodic and semantic memory.
Episodic memory is our memory for events and particular episodes;
semantic memory is, in simple terms, our general knowledge.
my knowledge that the capital of France is Paris is stored in semantic memory,
while my memory of a trip to Paris is an instance of an episodic memory.
• More example: Semantic or Episodic?
- dogs can’t fly.
- when you last rode a bike
- Christmas is on the 25th f December.
6
7. Big concept
The notion of meaning is closely bound to that of categorization.
A concept determines how things are related or categorized. It is a mental
representation of a category.
It enables us to group things together, so that instances of a category all have
something in common. Thus concepts somehow specify category membership.
All words have an underlying concept, but not all concepts are labeled by a
word.
For example, we do not have a special word for brown dogs. In English we have a word “dog”
that we can use about certain things in the world, but not about others. In principle we
could have a word, say “brog,” to refer to brown dogs. We do not have such a term, probably
because it is not a particularly useful one.
7
8. Cont…
Semantics concerns more than associations .
Words can be related in meaning without being associated (e.g., “yacht” and
“ship”), so any theory of word meaning cannot rely simply on word association.
Words with similar meanings tend to occur in similar contexts.
Lund, Burgess, and Atchley (1995) showed that semantically similar words
(e.g., “bed” and “table”) are interchangeable within a sentence; the resulting
sentence, while maybe pragmatically implausible, nevertheless makes sense.
- The child slept on the bed.
- The child slept on the table.
- The child slept in the cradle.
- *The child slept in the baby.
8
9. The denotation of a word
is its core, essential meaning.
Classic approaches to semantics
The connotations of a word
are all of its secondary implications, or
emotional or evaluative associations.
9
For example, the denotation of the word “dog” is its core meaning: it is the
relation between the word and the class of objects to which it can
refer. The connotations of “dog” might be “nice,” “frightening,” or “smelly.”
Put another way, people agree on the denotation, but the connotations
differ from person to person.
10. Consider the words “Hesperus” (Greek
for “The Evening Star”) and “Phosphorus”
(Greek for “The Morning Star”). They have
the same referent in our universe, namely
the planet Venus, but they have different
senses. The ancients did not know that
Hesperus and Phosphorus were the same
thing, so even though the words actually
refer to the same thing (the planet Venus),
the words have different senses (Johnson-
Laird, 1983). The sense of “Hesperus” is the
planet you can see in the evening sky, but
the sense of “Phosphorus” is the one in the
morning sky.
10
11. Semantic Network
One of the most influential of all processing approaches to meaning is based on the idea that
the meaning of a word is given by how it is embedded within a network of other meanings.
Some of the earliest theories of meaning, from those of Aristotle to those of the behaviorists,
viewed meaning as deriving from a word’s association.
From infancy, we are exposed to many episodes involving the word “dog.” For the behaviorists,
the meaning of the word “dog” was simply the sum of all our associations to the word: It
obtains its meaning by its place in a network of associations.
The meaning of “dog” might involve an association with “barks,” “four legs,” “furry,” and so on.
It soon became apparent that association in itself was insufficiently powerful to be able to
capture all aspects of meaning.
There is no structure in an associative network, with no relation between words, no hierarchy
of information, and no cognitive economy.
11
12. The Collins and Quillian senanic network model
A semantic network is particularly useful for representing information
about natural kind terms. These are words that denote naturally
occurring categories and their members—such as types of animal, or
metal, or precious stone. The scheme attributes fundamental
importance to their inherently hierarchical nature:
For example, a bald eagle is a type of eagle, an eagle is a type of bird of prey,
a bird of prey is a bird, and a bird is a type of animal.
12
13. Example of a hierarchical semantic network (based on Collins & Quillian, 1969)
13
15. The sentence verification task
One of the most commonly used tasks in early semantic
memory research was sentence verification. Participants are presented
with simple “facts” and have to press one button if the sentence is
true, another if it is false. The reaction time is an index of how
difficult the decision was. Collins and Quillian (1969) presented
participants with sentences such as:
- A robin is a robin.
- A robin is a bird.
- A robin is an animal.
- A robin is a fish.
15
16. Problems with the Collins and Quillian mode
First, clearly not all information is easily represented in hierarchical form.
What is the relation between “truth,” “justice,” and “law,” for example?
Second problem is that the materials in the sentence verification task
that appear to support the hierarchical model confound semantic distance
with what is called conjoint frequency. Such as The words “bird” and “robin”.
Conjoint frequency is a measure of how frequently two words co-occur. When
you control for conjoint frequency, the linear relation between semantic
distance and time is weakened (Conrad, 1972; Wilkins, 1971).
Third, the hierarchical model makes some incorrect predictions. For example:
- A cow is an animal.
- A cow is a mammal.
16
17. Revisions to the semantic network model
Collins and Loftus (1975) proposed a revision
of the model based on the idea of spreading
activation. The structure of the network
became more complex, with the links between
nodes varying in strength or distance.
(Example of a spreading activation semantic network. It
should be noted that two dimensions cannot do justice to
the necessary complexity of the network. Based on Collins
and Loftus (1975).)
17
18. Semantic features
The Semantic Features are 'categories' that allow us to classify the meaning of a word.
Semantic features work very well in some simple domains where there is a clear relation
between the terms.
We can take the idea of semantic features further, and represent the meanings of all words in
terms of combinations of as few semantic features as possible
They are usually represented with a headword and using the symbols "+' and' to say if the word
contains that feature.
There is no list of semantic features. We need the context to create one.
18
19. feature Father Mother Daughter Son
Human + + + +
Older + + - -
Female - + + -
19
Semantic features work very well in some simple domains where there is a clear relation
between the terms. They are usually represented with a headword and using the symbols "+'
and' to say if the word contains that feature. One such domain, much studied by
anthropologists, is that of kinship terms. A simple example is shown in the Table:
(Decomposition of kinship terms)
20. Early decompositional theories
Katz and Fodor (1963) showed how the meanings of sentences could
be derived by combining the semantic features of each individual word in the
sentence. It emphasized how we understand ambiguous
words. For example:
- The witches played around on the beach and kicked the ball.
- The witches put on their party frocks and went to the ball.
- ? The rock kicked the ball.
20
21. Feature-list theories and sentence verification
Rips et al. (1973) proposed that there are two types of semantic feature.
Defining features are essential to the underlying meaning of a word, and relate
to properties that things must have to be a member of that category .
(for example, a bird is living, it is feathered, lays eggs, and so forth).
Characteristic features are usually true of instances of a category, but are not
necessarily true (for example, most birds can fly, but penguins and ostriches cannot).
According to Rips et al., sentence verification involves making comparisons of
the feature lists representing the meaning of the words involved in two stages.
For this reason this particular approach is called the feature-comparison
theory.
21
22. Evaluation of decompositional theories
Hollan (1975) argued that it is impossible to devise an experiment to distinguish
between feature-list and semantic network theories because they are formally
equivalent, in that it is impossible to find a prediction that will distinguish
between them . Hence for all intents and purposes we can consider network
models to be a type of decompositional model.
decompositional theories have an intuitive appeal, and they make explicit
how we make inferences based on the meaning of words in the sentence
verification task.
it is difficult to construct decompositional representations for even some of the
most common words. Some categories do not have any obvious defining features
that are common to all their members. For example “ Game”
22
23. there are two important issues (McNamara &
Miller, 1989):
The first is whether we represent the
meanings of words in terms of features.
The other is whether we make use of
those features in comprehension.
when we see a word like “bachelor,” is the meaning
of the unmarried man sense of “bachelor” must
clearly contain features that correspond to
(+unmarried, +man), although these in turn might
summarize decomposition into yet more primitive
features.
23
Is semantic decomposition obligatory?
24. Family resemblance models
Many categories seem to be defined by a family resemblance
between their members rather than the specification of defining features
that all members must possess.
24
25. Prototype theories
A prototype is an average family member (Rosch, 1978).
Potential members of the category are identified by how closely they resemble the prototype
or category average.
The prototype is the “best example” of a concept, and is often a non-existent, composite
example. For example, a blackbird (or alternatively, American robin) is very close to being a
prototypical bird; it is of average size, has wings and feathers, can fly, and has average
features in every respect.
A prototype is a special type of schema. A schema is a frame for organizing knowledge that
can be structured as a series of slots plus fillers. A prototype is a schema with all the slots
filled in with average values. For example, the schema for “bird” comprises a series of slots
such as “can fly?” (“yes” for blackbird and robin, “no” for penguin and emu), “bill length”
(“short” for robin, “long” for curlew), and “leg Length” (“short” for robin, “long” for stork).
25
26. Basic levels
Rosch (1978) argued that a compromise between cognitive economy and maximum
informativeness results in a basic level of categorization that tends to be the
default level at which we categorize and think, unless there is particular reason to
do otherwise.
we use the basic level of “chairs,” rather than the lower level of “armchairs” or the
higher level of “furniture.” there is a basic level of categorization that is
particularly psychologically salient (Rosch et al., 1976).
Rosch et al. (1976) showed that basic levels have a number of advantages over
other categories. Participants can easily list most of the attributes of the basic
level; it is the level of description most likely to be spontaneously used by adults;
sentence verification time is faster for basic-level terms; and children typically
acquire the basic level first.
26
27. Problems with the prototype model
Hampton (1981) pointed out that not all types of concepts appear to have
prototypes: Abstract concepts in particular are difficult to fit into this
scheme. What does it mean, for example, to talk about the prototype for “truth”?
The prototype model does not explain why categories cohere.
Lakoff (1987) points to some examples of very complex concepts for which it is far
from obvious how there could be a prototype—the Australian Aboriginal language
Dyirbal has a coherent category of “women, fire, and dangerous things” marked by
the word “balan.”
27
28. Instance theories
We make semantic judgments by comparison with specific stored instances. This is
the instance approach (Komatsu, 1992), also called the exemplar theory.
There are different varieties of the instance approach, depending on how many
instances are stored, and on the quality of these instances.
The instance approach provides greater informational richness at the expense of
cognitive economy.
It is quite difficult to distinguish between prototype and instance-based theories.
Many of the phenomena explained by prototype theories can also be accounted for
by instance-based theories.
28
29. COMBINING CONCEPTS
Wisniewski and Love (1998) showed that in certain circumstances people
prefer to comprehend noun combinations on the basis of property relations.
High similarity between the constituents of a combination facilitates the
production of property relations.
People then look for a critical difference between them that can act as the
basis of the interpretation.
For example, consider “zebra horse.” “Zebra” and “horse” are close in
meaning, and the critical difference “has stripes” can easily be used to
generate the property relation “a horse with stripes.”
29
30. FIGURATIVE LANGUAGE
Humans make extensive use of non-literal or figurative language. In this we go
beyond the Literal meanings of the words involved, for humor, effect,
politeness, to play, to be creative—and for a mixture of these and other
reasons.
Metaphor: is one of the main types of figurative language. It is a special
type of conceptual combination, where we combine two concepts that are not
normally thought of as being related for some special effect. There are many
types of metaphor, depending on the relation between the words actually used
and the intended meaning. Here are a few examples:
- Vlad fought like a tiger. (Simile)
- Vlad exploded with fury. (Strict metaphor)
- All hands on deck. (Synecdoche)
30
31. Idiom: is also one of the main types of figurative language. It can be thought
of as frozen metaphors. Whereas we make metaphors up as we go along,
idioms have a fixed form and are in general use. The meaning of an idiom is
usually quite unrelated to the meaning of its component words.
Examples include “to kick the bucket” and “fly off the handle.”
Gibbs (1980), using reading times, found that participants take less time to
comprehend conventional uses of idioms than unconventional, literal uses,
suggesting that people analyze the idiomatic senses of expressions before
deriving the literal, unconventional interpretation.
Swinney and Cutler (1979) also found that people are as fast to understand
familiar idioms as they are comparable phrases used non-idiomatically. They
suggested that people store idioms like single lexical items.
31
32. THE NEUROSCIENCE OF SEMANTICS
Shallice (1988; see also Warrington & Cipolotti, 1996, and
Warrington & Shallice, 1979) discussed five criteria that could distinguish
problems associated with the loss of a representation from problems
of accessing it.
First, performance should be consistent across trials. If an item is
permanently lost, it should never be possible to access it. If an item is
available on some trials rather than on others, the difficulty must be one of
access.
32
33. Second, for both degraded stores and access disorders, it should be easier to
obtain the superordinate category than to name the item, because that
information is very strongly represented; but once the superordinate is
obtained, it will be very difficult to obtain any further information in a
degraded store.
Warrington (1975) found that superordinate information (e.g., that a lion is
an animal) may be preserved when more specific information is lost. She
proposed that the destruction of semantic memory occurs hierarchically, with
lower levels storing specific information being lost before higher levels storing
more general information.
33
34. Third, low-frequency items should be lost first. Low-frequency items should be
more susceptible to loss, whereas problems of access should affect all levels
equally.
Fourth, priming should no longer be effective, as an item that is lost obviously
cannot be primed.
Fifth, if the knowledge is lost then performance should be independent of the
presentation rate, whereas disturbances of access should be sensitive to the
rate of presentation of the material.
34
35. The structure of semantic memory: Evidence from
studies of dementia
Dementia is a general label for the widespread
decay of cognitive functioning, generally found
in old age. The ultimate causes of dementia are
unknown, although it is likely that both genetic
and environmental factors play some role, and it
is clear that there are several subtypes, the most
common of which is Alzheimer’s disease (AD).
In dementia, memory and semantic information
are particularly prone to disruption.
35
36. Semantic memory disturbances in dementia
There is a huge body of work indicating problems with
semantic processing in dementia.
People with dementia are often impaired on the category fluency task, where they have to
list as many members as possible of a particular category (e.g., Martin & Fedio, 1983).
They have difficulty listing attributes that are shared by all members of a category
(Martin & Fedio, 1983; Warrington, 1975).
They have difficulty in differentiating between items from the same semantic category
(Martin & Fedio, 1983).
They tend to classify items as being similar to different items more than controls do
(Chan et al., 1993a, 1993b).
They are also poor at judging the semantic coherence of simple statements:
For example, they are more likely to judge “The door is asleep” to be a sensible statement
than controls (Grossman, Mickanin, Robinson, & d’Esposito, 1996).
36
37. Difficulties with picture naming
People with dementia often have difficulty in naming things.
There is evidence that the semantic deficit is involved in picture naming.
Most of the naming errors in dementia involve the production
of semantic relatives of the target (e.g., Hodges, Salmon, & Butters, 1991).
The extent of the naming impairment is correlated with the extent of the more
general semantic difficulties (Diesfeldt, 1989).
Naming performance in dementia is sometimes affected by the semantic variable
of imageability.
With other types of neuropsychological damage, patients usually find high-
imageable items easier than low-imageable items.
37
38. Connectionist Approaches to Semantics
Connectionism has made an impact on semantic memory, just as it did in
earlier years on lower level processes such as word recognition.
This approach gives rise to the idea that semantic memory depends on
semantic microfeatures.
Note that this approach is not necessarily a competitor to other theories
such as prototypes; one instance of a category might cause one pattern
of activation across the semantic units, another instance will cause another
similar pattern, and so on.
38
39. Semantic microfeatures
A microfeature is an individual, active unit; the prefix “micro” emphasizes
that these units are involved in low-level processes rather than explicit
symbolic processing (Hinton, 1989), but there really isn’t much difference
between a feature and a microfeature.
Connectionist models suppose that human semantic memory is based on
microfeatures.
A semantic microfeature is really just a semantic feature, but the prefix
“micro” is added in computational modeling to emphasize their low-level nature.
39
40. Explaining language loss in people with Alzheimer’s disease: The
semantic microfeature loss hypothesis
What happens if a disease such as dementia results in the loss of semantic
microfeatures?
The effect will be to distort semantic space so that some semantic attractors might
be lost altogether, while others might become inaccessible on some tasks because of
the erosion of the boundaries of the attractor basins.
Damage to a subset of microfeatures will lead to a probabilistic decline in
performance.
Depending on the importance of the microfeature lost to a particular item in a
particular patient, the pattern of performance observed will vary from patient to
patient and from task to task.
Different tasks will give different results because they will provide differing amounts of
residual activation to the damaged system.
40