Greg Detre
Tuesday, November 12, 2002
Deb: how
straightforward was it to drop ThoughtTreasure into CycL?
grids are basically a set of assertions � when converted into CycL, he
just bundled all the coordinates into a single assertion
the hard
problem he�s trying to address is story understanding
the make
sense rating
there are weights at every level of processing (even down to the lexical
level, to signal archaic etc.)
do those weights have to be hard-coded too???
yes, they�re set manually
is there a hope that the mass of interactions between lots of different
weights, and at different levels, will just end up robust somehow???
Deb: what�s the resolution of the weights?
they�re just a decimal between 0 and 1
Erik would prefer it to be done by some sort of reasoning or automated
process eventually�
the assertions don�t have weights
ThoughtTreasure doesn�t have Cyc-like contexts
he started coding some philosopher belief systems, in the form �A
believes X�, but there�s no other major context mechanism
models
his methodology is: construct simulations of
the states and events described in the utterance
read: Johnson-Laird (1993), Human and machine thinking � agrees with the
model approach
the ThoughtTreasure approach is to use diverse representation schemes to
generate lucid representations
how would his grids cope with the table rotated
37deg???
presumably, it would simply try and map the real world story onto the
representation-simplification, rather than trying to accurately represent the
real world � but that does impose a ceiling on how well it can ever perform
(and there�s a quantum leap between these simplifications and representations
that can (potentially) scale to real-world complexity)
presumably, his hope is that you could insert a
lower level that could build complicated models that could be schematised into
ThoughtTreasure form
goals:
can either be binary achievements, or analog
understanding
agents:
why this approach?
convenience of modularisation
you can answer any question up to a certain level of granularity
similar to:
Karma � X-schemas???
grids:
can it do spatial reasoning with the grids???
distance, relative positions, line of sight/hearing, path-finding
Deb: if Jim and Bob are in an L-shaped room, can they see each other?
ThoughtTreasure will instantiate the room concretely � the answer will
depend on how it instantiates things
Deb: you could leave levels of Kuipers� spatial semantic hierarchy
unbound
problems he has: how do you dynamically instantiate
a grid
could you not circumvent most of these common
sense na� physics conundrums by interrogating a semi-realistic 3D physics
model???
well, how do you leave things unbound in such a model???
rather than doing a coarser grid, maybe you should use a
topology/containment approach (which he already has at the object-level)
Deb: he�s all for multiple representations, if you can move between them
automated +
semi-automated acquisition techniques:
e.g. adding grids using Open Mind
can planning + understanding agents be learned?
Brian: does
he have a systematic evaluation method?
similar approach to the information-extraction competitions
c. 40 baseline questions for story understanding, that you can ask after
more or less every sentence
e.g. what�s the character�s goal, what did he do, what�s his emotional
response etc.
if it
doesn�t know the answer, it produces nothing
you can ask
it more or less anything that it understands syntactically, and that it is able
to represent within its story system
however, it�s a very baby syntax, and a very
limited repertoire of encoded story-simulations
you also have a combinatoric problem, e.g. if you have five interacting
goals
Deb: is
there a notion of conceptual primitives (vs derived)?
he asked my question!!! you could list endless inferences like �you
sleep lying down�, but you wouldn�t need to if you had a physics/sleep model �
are you starting to see some primitives?
Vegas � the Manifesto for Lexicons:
concludes that conceptual primitives don�t work, and advocates a
conceptual primitive for every concept
his aim with ThoughtTreasure was to let the primitives emerge from the
mass
Deb: what
do words mean to ThoughtTreasure � what does a word like �sleep� mean, and what
bindings does it have?
sleep might have 5 inflections, linked to the sleep concept node, which
will be linked to a bunch of assertions
shrdlu � the lack of modularity of design leads to a lack of ways of
thinking of about it
open mind � comes closer to Pustejovsky�s aristotelian categories,
because it has the templates, but they�re pretty shallow
how did he
decide on the workings of emotional finite state automata???
Ortony�Collins model of emotions � it�s missing some, though they�re to
be found elsewhere in the ontology
given the
domains he�s chosen to focus on, is it at the stage where it could read and
learn from children�s stories???
it will produce a parse of a text � you can then try and generalise from
those, somehow (though it hasn�t been implemented yet)
so I spose the answer is, �kind of�
there�s an information-extraction feature that outputs NL information
into its own KR format
choice of
representation:
Lavecque � �lucid� (logic): for every predicate, for every possible
sequence of arguments, you have to know whether they�re true (given finite
universe of symbols)
completely unambiguous, corresponds to a single model in logic
e.g. a complete 3D model of something, or in logic the closed world
assumption
Push: lucid doesn�t have to be very concrete
but it does have to be fully-specifiable, right???
a lot of problems came up using non-lucid representations, e.g. in
logical AI (huge problems of theorem-proving combinatorial explosions blowing
up with non-lucid representation)
also, you have the problem with the frame-problem � for it to be a lucid
representation of something that involves time, you have to a have a full
specification of state for that entire time period
automated
augementation of you�re the representations he has
trying to build new planning agents from
corpus-based project (reading stories)
difficult parts: sequences, loops,
conditioned-gotos, goal, links to lexical entries
new
representations:
what about introducing a na� physics model?
currently, all of that�s implemented as procedural knowledge
but if you want to invert that procedural knowledge, to abduction (where
you know the conclusion, and part of the middle, and you want to fill in some
more assumptions etc.), he doesn�t know how
the declarative versions of the planning agents which he built to try
and solve this problem aren�t as rich
Deb: for abduction, can�t you just enumerate all the different state
paths that could have got you there?
combinatorial explosion
Hobbes � Tacitus � logical story understanding � incorporates a metric
for doing abduction to minimise the number of links
you could use Bayesian rules to narrow that search � it seems analogous
to the techniques used in speech recognition (again, it�s just nodes with
directed links)
Deb: does
the machine have a goal?
the goal is to find the highest make-sense rating though, isn�t it???
the goal is implicit in the mechanisms, isn�t it???
you need there to be more than one goal, and a choice of actions, in
order for it to be meaningfully a goal
Push: if the machine can measure whether it�s making progress, it can have
a goal
he wrote Daydreamer 15 years ago� multiple
goals, meta-goals, executive functions, connected to emotional data structures
Deb: fast
car vs fast road � modularity of concepts, modifying each other in different
ways � if you haven�t come across one combination before, is there any
mechanism for dealing with that
Pustejovsky is factoring the complexity, rather than just trying to
encode facts willy-nilly
Deb: likes
procedural knowledge
how do you
deal with contradiction???
e.g. animals often speak in children�s stories
Lenat, �12 dimensions of context space� (quite recent)
different from Cyc context
e.g. region, temporal, etc.
specify the context by twiddling those parameters
Push: he wouldn�t have been able to come up with that without just
ploughing through loads of assertions
rich, but
brittle???
what does
TT do that CyC doesn�t???
grids, scripts, different/simpler encoding, multi-lingual
how is it going to scale??? learn???
how does it deal with contexts???
see above � it doesn�t really
how flexible are the grids???
what do you hope to gain from being bi-lingual???
started out as a translation system
reminds you to stay general/cross-linguistic???
y, helped him find the appropriate level of generality
IBM funds this, right??? how???
is he happy for 1-D space???
how did he envisage investigating the space vs time cross-over???
is he not worried that the effects noted by Boroditsky�s experiments are particular to the human brain, and that we might need all sorts of idiosyncratic baggage to replicate them???
is there any reason to think that they would result from almost any space+time representation???
what does he hope to achieve by replicating them??? is he interested in showing some sort of biological plausibility???
how does he feel about using connectionist/hybrid representations???
which prepositions does he want to start with??? how many??? is there a sort of critical mass of prepositions you can use in your spatial communications???
I suggested near, then front/back:
�I could focus on the cross-over from representing
space to representing time. The best way that I�ve come up with of doing this
would be to have the objects approaching at different speeds. I think this
would still be possible within a one-dimensional world, which is why I chose
near/soon as my central proposition, with front/coming as potentially a
secondary proposition, to investigate.
I think that investigating the temporal effects
would be more likely to work, and might be potentially more interesting.
However, I suspect that the most interesting effects, like the ego- vs
time-moving representations, would require a pretty advanced cognitive
architecture to be seen. Indeed, they must be contingent to some degree on the
particular way in which the human brain represents space, time and movement,
and I can�t yet see how to strip away the extraneous factors to model them. If
we were investigating time in a two-dimensional world, then it might eventually
also be possible to consider horizontal vs vertical linguistic conceptions of
time as well.�
he said:
�I just looked at your proposal. First comment is
that I had asked for only one page -- that helps force you to focus more than
you have here. I think of your two final ideas, the second is better. We can
talk sometime next week about how to make it a practical project. Lera's
suggestion is to look at 'ahead' and 'behind' as a good starting point for
time-space conversion.�
is this is
the major NL book???
can you
really differentiate between general knowledge and knowledge of a specific
situation???
next week:
cognitive science
theory theory (Josh Tanenbaum)