Transcript Feldman, Jerome A. 2006. From Molecule to Metaphor. Cambridge

Feldman, Jerome A. 2006
From Molecule to
Metaphor Cambridge, MA:
Bradford MIT Books.
Short presentation by
Laura Janda and Tore
Nesset
1
The point of the book
• To promote linguistic models that
correspond to what we know about how
the brain functions
– Ex: Embodied Construction Grammar, ECG
• Models based on rules do not correspond
to what we know about the brain
– Alternative: “neural theory of language”
• Overview of the place of linguistics in
cognitive science
2
Advantages of the book
• A well balanced discussion with a lot of
information from various disciplines:
– neurology, computer science, and linguistics
• Does not just sum up results, but also
explains why they are important and the
relationships between them, presents
them as a story that has a plot
• Very clear and easy to read, with lots of
concrete examples
3
Disadvantages of the book
• Not a lot of news for people who are familiar with
cognitive linguistics
• The book is often focused only on research at
UC Berkeley
– but it is true that many important ideas were born in
Berkeley
•
•
•
•
•
Searle’s ”Chinese room”
Fillmore’s frames
Lakoff’s metaphor theory
Goldberg’s construction grammar
Narayanan’s, Bergen’s, and Grady’s work with metaphors
4
The structure of the book: From
Amoebas to Human language
• Begins with single cells, then goes to the brain,
mind, before describing increasingly complex
linguistic structures
• Important questions:
– What do we know about amoebas and how do they
categorize without rules?
– What do we know about neurons and the brain and to
what extent can they be compared with a computer?
– How can one build up linguistic models that take into
account how the brain works?
– What is meaning based on, and why are humans
much more clever at using language than other
beings?
5
The structure of this talk:
Main points
1. Category formation and modularity: Language
in the brain
2. Key concepts from cognitive linguistics and the
perspective of cognition
•
•
•
•
•
Prototype
Image schema
Frame
Embodiment
Metaphor
3. Embodied Construction Grammar
4. A question and some speculation to close with
6
Category formation and
modularity: Language in the
brain
PART I
7
Categorization
• “Categorization occurs whenever a lot of data
are boiled down to a few values.” (s. 96)
• Categorization is found in all living systems, not
just in language
• Amoebas categorize: food vs. non-food, danger
vs. non-danger
– This happens chemically, the outer membranes of the
amoeba react in different ways to food and non-food.
– These reactions cause changes in the shape of the
amoeba.
– It is possible to write rules for how the amoeba reacts,
but they don’t correspond to anything in the amoeba –
it categorizes without such rules.
8
More categorization
• Simple neural nets can categorize and react
without rules
– Knee-jerk reflex (just goes through the spine)
• It is possible to write a rule for it, but we know that the reflex
is not stored in the body as a rule, it is just a result of how
four neurons are connected to each other
• More complex systems, like dancing, can also be described
as rules, but this does not mean that they exist in the brain as
rules, nor that there is an autonomous ”dancing module”in
the brain (p. 279-80)
9
The brain
• We are used to comparing the brain to a computer, but there are
many important differences
– the brain has many more units than a computer
– the brain’s units (neurons) work much more slowly (106) than the units
in a computer
– the brian’s neurons are more connected to each other than the units in a
computer
– when the brain is used, the connections between the neurons are
changed (“neurons that fire together, wire together”)
– the brain is not autonomous – it is a part of the body and gets
information from the world and reacts to that information
– the brain has ”mirror neurons” that fire when a person does something
or when a persons sees that someone is doing that something or when
a person thinks about (simulates) doing that something
– mirror neurons give human beings a greater capacity to imitate
– only humans can simulate non-actual situations and therefore
understand things from different perspectives (understand what another
person is thinking)
10
Linguists ignore what we know
about the brain
• They claim that neurology is not sufficiently dveloped, so it
is not necessary to take it into account
– “neuroscience is not nearly developed enough to be taken
seriously” (Chomsky 2003)
• They assume that a grammar is a collection of formal rules
• They are interested only in compentence, not in
performance (and thus they can ignore corpus data)
• They assume that language is an autonomous module in
the brain
• Thus they make themselves independent of other
disciplines (p. 273)
• Feldman thinks this is irresponsible (p. 151)
11
Trying to provide a balance
• Feldman doesn’t just criticize Chomsky and the
generativists, he claims that all kinds of linguists are too
focused on isolated grammatical problems without
appreciating the role of meaning and language use (p.
281-2)
• Nature vs. Nurture: Formalists see only Nature,
functionalists see only Nurture, but
– it is a fact that the brain has a certain structure from the
beginning
– it is a fact that the brain is plastic and is always changing
• That which genes specify is perhaps only a unique
capacity to learn which is not further specialized for
language (p. 272).
12
What is learning?
• Both Nature and Nurture are cumulative and
influence each other
– Nature: genes are “expressed”, but all cells have the
same genes, and the parts of them that are activated
change over time
– Nurture: Leaning isn’t just a matter of adding
something to an already existing system; leaning
changes the system itself
• Learning is a strengthening of connections
between neurons.
• Learning creates categories – the world exists,
but there is more than one way of categorizing it
13
Key concepts
PART II
14
Prototypes
• Categories can be organized around
various types of prototypes:
– Reference point (100, 1000, circa)
– Scalar prototype (standard of measurement)
– Typical member
– Ideal member
– Salient example
• Prototypes er compatible with what we
know about the brain
15
Prototypes (cont’d.)
• Neural networks have weighted
connections and various grades of ”firing”
(p. 97).
– Gradual categorization in relation to a
prototype
• Neural systems have thresholds which
have to be exceeded before they will fire
at all:
– Either-or categorization (classical categories)
16
Image Schemas
• Talmy – image schemas show how language
organizes spatial relations, with trajectors and
orientation points
• Various types:
– Topological (container, path, etc.)
– Orientational (in relation to the body, for example, in
front of, behind)
– Force dynamic (against, etc.)
• Based on genetic inheritance and universal
human experience (p. 136)
• Feldman shows that image schemas can be
described as feature-value structures, but
doesn’t tell us exactly how the schemas are
related to neural networks.
17
English into combines two image schemas:
Container
Source-path-goal
inside
source
outside
path
boundary
goal
trajector
Meaning is a collection of relations among
image schemas (p. 284)
18
Frames
• Fillmore’s frames show how knowledge is
organized and how words/concepts are
connected to each other
– The “commercial event frame” is activated to
understand words like ”buy” and ”sell”
• Frames can be composed of scenarios
(phases):
– Initial state
– Exchange of money for goods
– Resulting state
• Frames give us the possibility to describe series
of events with characteristic elements (like buys)
and options for filling in slots (like John)
19
Frames (cont’d.)
• Frames are composed of simpler concepts
which in turn boil down to image schemas.
• Therefore frames are based in neural
networks.
• Frames are culturally conditions and thus
not universal.
• Image schemas can however be universal:
“universal bodily based representations of
experience” (p. 136).
20
Embodiment
• Used in two related meanings in cognitive linguistics:
– The meaning of a word is grounded in bodily experience with
situations where the word is used, what a human being does
with the relevant things or events
– The meaning of a word is directly anchored in neural networks in
the brain
• When small children play, they put things in containers
and this establishes an image schema (container) which
in turn serves as the basis for the meaning of words.
• The embodied basis for meaning also means that a
person can react to actual situations (the Chinese room
or a computer cannot react to questions about what color
they are seeing now, or what they should do if the
building is on fire)
21
Embodiment and the brain:
The knee-jerk reflex
• The same neural structures are active when (p.
4f.)
–
–
–
–
The doctor hits you on the knee
The doctor asks you to kick
You see someone kick
You hear about someone kicking
• Mirror neurons are activated during simulation
• Understanding/meaning arises through
simulation
• In this way linguistic meaning is directly
connected to neural networks
• Neural networks are very relevant for linguistics 22
Embodiment, simulation and
understanding
• Simulation semantics – understanding is
the capacity to simulate a narration
• The capacity to carry out an action on the
basis of linguistic description – this is the
key to understanding
• Computers and programs lack this
capacity
23
Metaphor
• How to link abstract thought with embodied experience?
• “Essentially all of our cultural, abstract, and theoretical
concepts derive their meanings by mapping through
metaphor, to the embodied experiential concepts we
explored in earlier chapters” (p. 199)
• Grady’s 1996 primary metaphors link subjective
evaluations and sensory-motor experiences:
–
–
–
–
–
Affection is warmth:
Subjective: Affection
Sensory-motor: Temperature
Example:
“They greeted me warmly”
Experience: A child feels warmth when held by a parent
• Metaphors are a normal consequence of associative
learning by means of neural networks: coactivation of
neurons (“Neurons that fire together, wire together” p.
201)
24
Metaphor (cont’d.)
• C. Johnson shows how metaphors are learned:
– First a child learns a literal meaning: See Daddy
– Then the child learns “conflations”: See what I spilled
– Then the child learns metaphorical meaning: See
what I mean
– “Conflation” involves coactivation.
• Complex metaphors involve the activation of several
metaphors at the same time, thus involving more
comprehensive neural networks.
• Narayanan 1997 developed a data program that
understands metaphors in a narrative, e.g., France fell
into a recession
– recession is a metaphorical hole
– it is linked to economic concepts
– the program understands that France wasn’t in the
hole before it fell, that it is still in the hole, and that it
25
couldn’t control its movement.
Embodied Construction
Grammar
PART III
26
False assumptions
1. A grammar is a collection of abstract rules
2. Formal grammatical rules are expressed in the brain
3. A grammar is independent of all other structures in the
brain
4. Genes provide specific grammatical information
•
2, 3, and 4 are very unlikely given what we know about the brain
and genes
5. Children do not get enough input to build a grammar
–
–
But we know that children get lots of input, that all input is in
context, and that the input doesn’t come in a random order
There is no “poverty of stimulus”, but an “opulence of substrate”,
because a child has a lot of conceptual and embodied
experience, plus support from other people (p. 318-19)
6. Each word has a number of set meanings, meanings are
in the words, and grammatical rules are abstract and
meaningless
–
But we know that the meaning of a word is affected by context,
27
by ongoing perceptions and associations
An embodied neural theory of language
• It is important to find a way to write a grammar that
corresponds to what we know about the brain and
neurons
• Embodied Construction Grammar (ECG) develops
a formal notation for cognitive linguistics (p. 297)
• The construction is the basic unit of speach =
form + meaning
– NB. Langacker’s symbolic structure, which
Feldman does not cite
– Meaning can come from larger structures than
morphemes (p. 298)
28
An embodied neural theory of
language (cont’d.)
• Embodied
– Grounded in embodied schemas (p.
289)
• Four basic formal structures
– Schemas
– Constructions
– Maps (which facilitate metaphor)
– Mental spaces (indirect speech, etc.)
29
A question and some
speculation to close with
PART IV
30
How did language develop?
• We don’t have any fossil languages and
language changes too fast – we don’t
know anything about an earlier or more
primitive version of language
• Chomsky, Hauser and Fitch (2002)
conclude that language (i.e. grammar) is
the result of just one big genetic mutation
• Pinker and Jackendoff believe that
language developed gradually
31
How did language develop?
• But if one believes that language is just the result of a
bigger brain with better learning capacity, there isn’t so
much to explain
• According to the neural theory of language, the capacity
to simulate played a major role in the development of
language
• One extra step: displacement – the capacity to simulate
and think about solutions which are not connected to the
here and now
• All mammals show displacement when they dream, but
only humans can do this willfully, when they are not
asleep/dreaming
• There is a relatively small step from animals’
uncontrolled displacement to humans’ controlled
displacement (p. 328)
32
Let’s sum up!
33
Feldman’s theory spans
• The interaction of amoebas
• The interaction of neurons
• Neural networks in the brain
• Literal meaning
• Metaphorical meaning
34
Feldman shows that
• Cognitive linguistics builds upon and is
congruent with what we know about the
brain.
• It is possible to formalize cognitive
linguistics via Embodied Construction
Grammar.
35