Greg Detre
@16 on Wednesday, October 30, 2002
Manchester
CHILDES corpus (Lieven, Pine, Rowland & Theakston) � CMU, transcripts +
audio recordings
Deb chose: 12 children recorded once a week, from 2-3 yrs age and
counted the words � these were the most common:
I
no
yeah
there
that
it
a
the
oh
one
on
in
and
you
this
go
want
to
mummy
got
my
it is
yes
do
mmhm
not
get
me
here
is
look
put
that is
what
some
now
have
down
car (41)
don�t
our
um
like
up
egoistic seems to come first
�car� (41st) is the first tangible noun
lots of deictics, pronouns
lots of cognitive states (aah, oh, ow)
synset =
set of words that can be used interchangeably
does Wordnet have �entailment� relations???
motivations: systematic polysemy
e.g.
fast car, fast road, fast food (speed it passes through you)
good food, good tree, good car
qualia structure
constitutive: what its made of (parts, substances)
formal: is-a
telic: direct + indirect function
agentive: origin
processes: type construction, type coercion
how would you fit �good tree� into the qualia structure???
Filmore
worth looking at if you�re interested in lexical semantics
circularity exists whenever you have a networked structure � alternatively, you end up with dangling nodes
that�s ok for people, because we can ground the (circular) concepts in the world
Toko�s fucking cool
you tell him �look, red cup� when he looks at something, and he can repeat it back to you and later search for that object on a table
cross-modal learning algorithm
histogram of all the possible pairwise tangent angles for a given object � rotation-invariant, and captures something holistic about the object
convex hull??? I think this just draws a perimeter all the way around
showed a 2D scene filled with coloured rectangles
3 layer grammar
phrase order
word class grammar
visual grounding of words
why voice reco??? resynthesises pieces of speech from the training input � sounds really realistic � generates grammatical sounding definite descriptions
wow
which paper??? he says it�s recent
Newt points at the object that you describes
why does it take so long??? is it because the computational expense???
Ki-yu(sp???) � find the green bag � which one � the green one
investigating lexical structure � if the robot had a dictionary, what would an entry that would make sense to it contain?
there has to be some picture of a cup that it can match to, but what else?
lexical semantics or world knowledge? his answer is �yes� � the closer he looks, the closer they seem
language is parasitic on non-linguistic structure
conceptual structures/relations are sometimes expressible through language (lexicalised or grammaticalised)
language has some distinct/arbitrary/autonomous aspects (e.g. phonology, syntax maps (e.g. word order relating to conceptual ordering � man bites dog vs dog bites man))
language as a lens/spotlight on underlying conceptual structure (hooks in to the underlying system, and provides entry points to under the hood)
phonology � how to say/recognise it
argument maps: semantic roles � syntactic roles
form � how it appears
behaviour � how it acts
interactive behaviour � how to react
procedural � how to do it
constitutive � what it�s made of
of specific instances
(all of the nature properties apply to types � this helps distinguish tokens)
what it is (used for, used to do)
model of how it works
how it came to be
the idea is, that if we want to build a dictionary for a robot, each word will have entries for these placeholders
discusses the schema for picking up an object at different levels
from grounding to object:
region � obj
what does the little red arrow that goes from grounding to objective do???
plug the lexical structure into the grounding machinery
perceptual-motor schemas � Ripley � rich interleaving of what you do and what you feel
e.g. affordance properties (stuck, missing etc.) � see children early vocabulary
3 systems
grounded
objective
intentional
linked by LeapOfFaith processes � these involve strong inferences
inter-connected systems instead of layers
who builds the robot hardware in his lab???