The Neural Theory of Language Project (How Can the Brain Compute Concepts?)
[Jan. 16 1998 - CS200]
How do human neural systems learn the specific kinds of concepts
that occur in natural languages and the language that expresses those
This is the question taken up in the Neural Theory of Language
(NTL) research group at the International Computer Science Institute at
Berkeley, a collaboration since the late 1980's between Jerome Feldman,
George Lakoff, and their students. At the heart of the modeling effort is
Feldman's idea of structured connectionism, which can be used to model
highly specific brain structures. The central enterprise of the group is to
provide neural models of embodied cognition, especially the acquisition and
use of language and thought as described in cognitive linguistics.
This talk will be an overview of three modeling efforts done in
dissertations by students in the group: Terry Regier's thesis on the
acquisition of spatial relations concepts and terms. David Bailey's thesis
on the learning of verbs of hand motion. And Srini Narayanan's thesis on
how aspect arises in neural control systems, on how metaphor is modeled,
and on how the same neural control system that can control body movements
can do abstract reasoning.
Since there is an enormous gap between the physical brain
and the level of human concepts and language, the NTL group has developed a
paradigm for ultimately bridging that gap in a small set of precise steps,
using research methodologies already in place within cognitive science.
Level 1. Cognitive Science and Cognitive Linguistics
Level 2. Neurally Reducible Conventional Computational Models
Level 3. Structured Connectionist Models
Level 4. Computational Neuroscience
Level 5. Neuroscience
The link between Level 4 and Level 5 is given by computational
neuroscience, which models the brain as if it were "circuitry," with axons
and dendrites seen as "connections," activation and inhibition as positive
and negative numerical values, and so on.
Since brain circuitry is enormously complicated, structured
connectionist models seek a simplified representation of such circuitry, in
which equivalent computations are carried out by neural circuitry of
minimal complexity. Thus, the link between Level 3 and Level 4 is one of
simplification: the minimal structured circuitry that will do the same job
as models of the actual "brain circuitry."
Sometimes the relation between the level of analysis in cognitive
science and cognitive linguistics and the structured connectionist level
can be given directly, as in the model by Regier. But even the simplest
neural models that carry out a complex task can be so complicated that an
intermediate level of representation is helpful in doing the modeling. In
the NTL paradigm, conventional computational systems from computer science
are sought out that have the basic properties of neural systems: parallel
operation, distributed control (no internal clock or centralized
controller), ability to react quickly and effectively to changing contexts,
resource dependency, ability to learn via statistical correlation, and so.
Such conventional computational mechanisms can form a link between level 2
and level 3. They are only used when it is known how to map them directly
to structured connectionist neural models.Members of the group have
constructued novel forms of conventional-style computational along these
guidelines. Only neurally reducible computational models are used. Very few
conventional computational systems satisfy all the requirements for neural
Such models help us to get a precise handle on how conceptual
structure is embodied, esopecially for the very specific, detailed
conceptual structures that cognitive linguists have uncovered.
more information on this research: The Neural Theory of Language Project
A Dynamic Model of Metaphoric Reasoning about Events
[Jan. 23 1998 - CS200]
This talk describes a new computational model for metaphoric reasoning
about event structure. A novel feature of the model is an "active"
representation of embodied motion and manipulation verbs (such as
walk, push, slide, slip) that is inspired by what is known about
high-level sensory-motor control and satisfies general constraints on
modeling neural activity. As a result, we are able to use a single
representation for controlling actions in a dynamic environment and
for context-sensitive, simulative inference in language understanding.
Monitoring and control parameters abstracted from the basic model
provide semantic grounding for interpreting aspectual expressions and
seem to offer elegant ways to solve the vexing linguistic problem of
aspectual composition. The implemented model is able to use metaphoric
projections of motion verbs to infer in real-time important features
of abstract plans and events; potentially explaining the frequent use
of motion and manipulation terms in discourse about abstract
plans. Results of applying our model on discourse fragments from
newspaper stories in international economics show that crucial facts
about abstract plans, goals, resources and intent are communicated by
projections from embodied concepts.
DIRECTLY RELEVANT PAPERS
Narayanan, Srini (1996). ´´Embodiment in Language Understanding:
the semantics of causal narratives''. AAAI Symposium on Embodied
Cognition and Action, Nov. 9-11, Cambridge Mass. Technical Report,
FS-96-02, AAAI Press, Menlo Park, CA.
Narayanan, Srini (1997). ´´Talking The Talk Is Like Walking the Walk:
Computational Model of Verbal Aspect''. Proceedings of the Nineteenth
Annual Meeting of the Cognitive Science Society Aug 9-11, Stanford:
Stanford University Press, 1997.
Narayanan, Srini (1997). ´´KARMA: Knowledge-Based Active
Representations For Metaphor and Aspect''. Ph.D. Dissertation,
Computer Science Division, University of California, Berkeley, 1997.
NTL PROJECT OVERVIEW PAPERS
Feldman, J, G. Lakoff, D. Bailey, S. Narayanan, T. Regier, A. Stolcke
(1996). ´´Lzero: The First Five Years''. Artificial Intelligence
Review, v10 103-129, April 1996.
Bailey, D., J. Feldman, S. Narayanan, and G. Lakoff (1997).
´´Embodied Lexical Development'', Proceedings of the Nineteenth
Meeting of the Cognitive Science Society, Aug 9-11, Stanford: Stanford
University Press, 1997.
postscript versions of readings (ICSI)
Conceptual Integration Networks
[Jan. 30 1998 - CS200]
Conceptual integration is on a par with other high level cognitive processes
such as analogy, induction, recursion, reference-point organization, conceptual
metaphor, and it interacts with all of them. It has systematic dynamic mapping
properties, obeys optimality constraints, and operates, with the same
structural characteristics, in a wide variety of cognitive phenomena, such as
synchronic and diachronic meaning formation, problem solving, action and
design, scientific evolution, humor, counterfactual reasoning. The study of
this cognitive operation sheds light on creative processes of conceptual change
with emergent structure in short term everyday human interaction, as well as in
long term evolution of thought.
Conceptual integration networks are constructed non- compositionally in
accordance with optimality principles that may compete with each other. Six
guiding principles have been proposed (integration, web, unpacking, topology,
good reason, and metonymy tightening). Emergent properties of conceptual
blends. which account for many types of creativity, are constrained by such
When viewed in terms of conceptual integration networks, superficially quite
different types of meaning construction (e.g. Fregean predicate structure,
counterfactuals, frame blends, metaphors) all fall on a single continuum.
Moreover, although the meanings are not truth-conditionally compositional, the
mapping schemes are compositional.
This reveals an underlying unity of meaning construction processes that cuts
across the traditional distinctions between core logical semantics, pragmatics,
figurative language, and rhetoric. The research ties in with work in cognitive
science on analogy, metaphor, framing, and grammar, and extremely interesting
applications have been suggested in music, literature, sign language, political
science, and mathematics.
Integration is commonly presented as a purely mental phenomenon, but as Ed
Hutchins has shown, we often make use of structure available in the world, or
culturally created, in order to anchor blends. Novel technology and design
typically involve integration, as in computer interfaces. More subtle is the
kind of grounding found in language use, and especially in sign language.
Scott Liddell has shown how meaning in ASL gets constructed through on-line
blending. Here, integration with material anchors is part of everyday language
use, with anchors provided either by the real environment (including the speech
participants), or by the construction of a surrogate space (with invisible
surrogates added to the real environment).
Uniform principles of integration operate both at high levels of scientific,
artistic, and literary thought and at lower levels of everyday thinking,
talking, and understanding. This calls into question the sharp distinctions
of nature often made between ordinary syntax and meaning, problem-solving and
reasoning, scientific discovery, artistic and literary creation.
Coulson, Seana. 1997. Semantic Leaps: Frame-Shifting and Conceptual
Blending. UCSD Ph. D. dissertation.
Coulson, S. and G. Fauconnier. (submitted). Fake Guns and Stone Lions:
Conceptual Blending and Privative Adjectives. [for Conceptual Structure,
Discourse, and Language, III]
Fauconnier, G. & Turner, M. (in press). Conceptual Integration Networks.
Liddell, Scott K. 1995. Real, Surrogate and Token Space: grammatical
consequences in ASL. In K. Emmorey and J. Reilly (eds.), Language,
Gesture,and Space. Hillsdale, NJ: Lawrence Erlbaum Associates,
Liddell, Scott. 1997. Grounded Blends, Gesture and Meaning in American Sign
Language Discourse. Fifth International Cognitive Linguistics Conference.
Mandelblit, Nili. 1997. Creativity and Schematicity in Grammar and
Translation: The Cognitive Mechanisms of Blending. Doctoral dissertation,
UCSD Cognitive Science.
Robert, A. (in press) Blending in Mathematical Proofs. Conceptual
Discourse, Structure, and Language II. Ed. Jean-Pierre Koenig. Stanford:
Center for the Study of Language and Information
Sweetser, Eve. 1997. "Mental Spaces and Cognitive Linguistics: A Cognitively
Realistic Approach to Compositionality." Fifth International Cognitive
Linguistics Conference. 1997.
Turner, Mark. 1996a. Conceptual Blending and Counterfactual Argument in the
Social and Behavioral Sciences in Philip Tetlock and Aaron Belkin, editors,
Counterfactual Thought Experiments in World Politics (Princeton:
Princeton University Press).
Zbikowski, Lawrence. 1997. "Conceptual blending and song." Manuscript.
University of Chicago.
more on blending and conceptual integration
Material anchors for conceptual blends
[Feb. 6 1998 - CS200]
An interesting class of cognitive processes utilize mappings between
conceptual spaces and physical spaces. This talk considers a number
of examples in which physical structures act as material anchors for
complex assemblages of conceptual relations. In these processes, the
spatial organization of material anchors plays a variety of roles
including support for conceptual blending, the substitution of robust
perceptual processes for conceptual processes, and overcoming the
limits of short-term memory. The examples range from everyday devices
such an analog wristwatch to specialized tools such as aircraft
instruments and the complex imagery used by micronesian navigators.
The two readings are a paper I wrote for the journal
Cognitive Science titled "How a cockpit remembers its speeds" and the
second chapter of my recent book "Cognition in the Wild." I intend
the cockpit speeds paper to be background for a discussion of the
airspeed indicator as a material anchor for a blend of space and speed
to yield a blended speed-space in which perceptual operations on
spatial relations of elements of the display stand in for conceptual
operations concerning speed. This is an emergent property of the blend
realized in material form.
The chapter from Cognition in the Wild contains two examples that I
will discuss in the talk. One is the system of micronesian navigation
in which navigators in the western Pacific ocean use complex
imagery that superimposes several conceptual frames into a single
blended space. The other is the way Medieval navigators in northern
Europe figured out how to look at a compass rose and see the tides
Dept. of Psychology
University of California, Los Angeles
Relational Reasoning in a Physical Symbol System
[Feb. 13 1998 - CS200]
Is the human mind a physical symbol system, and if so what does this
imply for theories of reasoning and learning? I will argue that the
mind is symbolic in a way that requires dynamic variable binding over
distributed representations of meaning. Neither PDP-style
connectionist models nor traditional symbolic models satisfy these
twin requirements. These requirements are met by symbolic
connectionism, an approach I will illustrate using the LISA model of
John Hummel and Keith Holyoak. 1997. Distributed Representations
of Structure: A Theory of Analogical Access and Mapping.
Psychological Review Vol. 104, no. 3, 427-466.
Keith Holyoak and John Hummel. in press. The Proper Treatment of Symbols in a Connectionist Architecture. In E. Deitrich & A. Markman (Eds.), Cognitive dynamics: Conceptual change in humans and machines. Cambridge, MA: MIT Press.
John Hummel and Keith Holyoak. 1996. LISA: A Computational
Model of Analogical Inference and Schema Induction.
18th CogSci Society (352-357).
Review of 'Mental Leaps' (by Doug Hofstadter). AI Magazine. 1995.
Dept. of Psychology
University of Arizona
To be real:
Representation and misrepresentation in models of concept combination
[Feb. 20 1998 - CS200]
I review a number of models of the concept combination coded by simple
noun phrases such as "brown cow," "chocolate bunny," and "fake ID". While
many models seem to be able to explain how people combine concepts to yield
"brown cow," combinations such as "chocolate bunny" and "fake ID" have proven
more challenging. I suggest that these limitations result from three things:
first, the assumption that concepts are static structures in long term
memory; second, commitment to the idea of a conceptual core; and third, the
idea that concept combination is the algorithmic combination of objective
features. In contrast, I suggest that concepts are dynamically assembled in
response to contextual needs, and that relational mapping operations are
fundamental to concept combination. I present a variety of data which support
this position and show how a comprehensive theory of concept combination can
be developed from conceptual integration theory (Fauconnier & Turner, in
press). Overall, these findings suggest that the mapping operations
underlying concept combination are dependent on (i) causal information, (ii)
relational information, and, even, (iii) human theory of mind.
Coulson, S. 1996. Menendez Brothers Virus: Blended Spaces and Internet
Humor. In A. Goldberg, ed., Conceptual Structure, Discourse, and Language.
Coulson, S. and G. Fauconnier. ms. Fake Guns and Stone Lions: Conceptual
Blending and Privative Adjectives.
Franks, B. 1995. Sense Generation: A "quasi-classical" approach to
concepts and concept combination. Cognitive Science 19:441-505.
Arthur Markman (Columbia)
Cognitive models of analogy, similarity, and metaphor
[Feb. 27 1998 - CS200]
There has been a remarkable degree of convergence among researchers in the
last ten years about how to think about the process of analogical mapping in
humans. In particular, the field has generally agreed that analogical
comparisons involve finding common relational structures across domains. Based
on the shared grounding assumptions, a number of computational models of
analogy have been developed including Falkenhainer, Forbus, & Gentner's
SME, Hummel & Holyoak's LISA, Keane's IAM, Holyoak & Thagard's ACME and
Hofstadter and Mitchell's COPYCAT. These models differ from each other in the
particular constraints on analogical mapping they value and on the
architectures they use to create analogical matches.
The shared foundational assumptions among researchers have allowed the
research community to define important disagreements about aspects of the
mapping process. We will begin with a discussion of the points of agreement
among researchers. Then, we will focus on key areas of disagreement including:
(1) the importance of structure in analogical mapping, (2) the role of semantic
similarity in mapping, (3) the development of representations, and (4) the
nature of representational change. In order to examine these disagreements we
will explore the computational models and also discuss empirical evidence from
studies of analogy, similarity and metaphor processing in humans.
Forbus, K. D., Gentner, D., & Law, K. (1995). MAC/FAC: A model of
similarity-based retrieval. Cognitive Science, 19(2), 141-205.
Forbus, K. D., Gentner, D., Markman, A. B., & Ferguson, R. W. (in press).
Analogy just looks like high level perception: Why a domain-general approach
to analogical mapping is right. Journal of Experimental and Theoretical
Gentner, D., & Markman, A. B. (1997). Structural alignment in analogy and
similarity. American Psychologist, 52(1), 45-56.
Gentner, D. & Wolff, P. (in press). Metaphor and knowledge change. To
appear in E. Dietrich, & A. B. Markman (Eds.) Cognitive dynamics.
Cambridge, MA: The MIT Press.
Markman, A. B. (in press). Constraints on analogical inference. Cognitive
Additional (background) readings
Gentner, D., & Wolff, P. (1997). Alignment in the processing of metaphor.
Journal of Memory and Language, 37, 331-355. [Empirical evidence about metaphor
Hummel, J. E., & Holyoak, K. J. (1997). Distributed representations of
structure: A theory of analogical access and mapping. Psychological Review,
104(3), 427-466. [An introduction to LISA]
Keane, M. T. Ledgeway, T., & Duff, S. (1994). Constraints on analogical
mapping: A comparison of three models. Cognitive Science, 18, 387-438. [An
introduction to IAM, and a comparison of IAM, SME and ACME.]
Dept. of Psychology
University of California Santa Cruz
Embodied Metaphor In Language, Thought, And Culture
[March 6 1998 - CS200]
My talk will focus on how ordinary and poetic language
reveals significant insights into the embodied
character of metaphor in language, thought, and
culture. I will describe how systematic analyses
of different linguistic patterns (e.g., in word
meanings, conventional expressions, and novel,
poetic discourse) suggest the importance of
embodiment in how people make sense of different
aspects of their experience, and provide partial
motivation for the ways people speak about these
experiences. Special attention will be given to
looking at how recurring patterns of embodied
experience give rise to "image schemas" that
can be metaphorically elaborated upon to help
structure many basic abstract concepts.
I will describe different aspects of my research
program looking at how embodied metaphor influences
people's interpretations and on-line comprehension
of word meanings, conventional expressions, and
poetry. Following this, I suggest several ways that
this empirical research constrains theories of
meaning and concepts, particularly in regard
to acknowledging the necessity of metaphor and
embodiment in both language and thought.
Finally, I will argue that cognitive scientists
must be very careful NOT to assume that just
because metaphor is indeed conceptual (and not
only a linguistic device), it must necessarily
be viewed as another kind of internal mental
representation. An important part of my present
research strategy is to explore the ways that
metaphor is "off-loaded" into the cultural world.
*Gibbs, R. (1996). Why many concepts
"Cognition" 61, 309-319.
Gibbs, R. (1994). "The poetics of mind: Figurative
thought, language, and understanding." New York:
Cambridge University Press. (see chapters 1, *4, and 6).
Gibbs, R., Beitel, D., Harrington, M., & Sanders, P.
(1995). Taking a stand on the meanings of "stand":
Bodily experience as motivation for polysemy.
"Journal of Semantics, " 11, 231-251.
*Gibbs, R. (1997). How language reveals the embodied nature
of creative cognition. In T. Ward, S. Smith, and J.
Vaid (Eds.), "Creative thought." Washington, DC:
American Psychological Association.
Gibbs, R., Bogdonovich, J., Sykes, J., & Barr, D. (1997).
Metaphor in idiom comprehension. "Journal of
Memory and Language," 37, 141-154.
Center for Research on Concepts and Cognition
At the Core of Human Creativity...
[March 13 1998 - CS200]
Using examples from the worlds of art and literature, I will attempt to
describe how a polished creation emerges out of a broth of many diverse
cognitive acts: internalization, essence-seeking, quality judgment, flexing,
exporting, adapting, uniformity-imposition, symmetry-favoring, echo-insertion,
tracks-covering, and others.
These kinds of mechanisms -- as well as a manner of allowing them all to
coexist and to exert their influences in parallel on developing creations --
have been implemented, albeit in a rudimentary manner, in computer programs
developed by my associates and myself over the past couple of decades. All of
them operate in micro-worlds that anyone can understand and relate to, and in
which, by great fortune, the key cognitive issues also emerge with maximal
The work is ongoing, and one important avenue of progress is the insertion of
greater self-awareness into such systems, so that they are more "conscious" of
what they are doing, and thus what they do is more genuinely "under their
The focus of our investigations is the nature of esthetic creation, with all
the vagueness and subjectivity that that inevitably entails. Our goal might
be characterized as that of describing in somewhat objective terms the
underpinnings of the seemingly intangible and murky yet very real and powerful
subjective impulses that always lie behind any serious creative effort.
Chapters 4, 5, 6, 7 of
Hofstadter, Douglas R. and the Fluid Analogies Research Group. 1995. Fluid
Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms
of Thought. Basic Books.