Ben and Stephen,
AFAIK your focus - and the universal focus - in this debate on how and
whether language can be symbolically/logically interpreted - is on
*individual words and sentences.* A natural place to start. But you
can't
stop there - because the problems, I suggest, (hard as they already are),
only seriously begin when you try to interpret *passages* - series of
sentences from texts - and connect one sentence with another. Take:
"John sat down in the carriage. His grim reflection stared at him through
the window. A whistle blew. The train started shuddering into motion, and
slowly gathered pace. He was putting Brighton behind him for good. And
just
then the conductor popped his head through the door."
I imagine you can pose the interpretative questions yourself. How do you
connect any one sentence with any other here? Where is the whistle
blowing?
Where is the train moving? Inside the carriage or outside? Is the
carriage inside or outside or where in relation to the moving train? Was
he
putting Brighton *physically* behind him like a cushion? Did the
conductor
break his head? etc. etc.
The point is - in reading passages, in order to connect up sentences, you
have to do a massive amount of *reading between the lines* . In doing
that,
you have to reconstruct the world or parts of the world, being referred
to,
from your brain's own models of that world.. (To understand the above
passage, for example, you employ a very complex model of train travel).
And this will apply to all kinds of passages - to arguments as well as
stories. (Try understanding Ben's argument below).
How does Stephen or YKY or anyone else propose to "read between the
lines"?
And what are the basic "world models", "scripts", "frames" etc etc. that
you
think sufficient to apply in understanding any set of texts, even a
relatively specialised set?
(Has anyone seriously *tried* understanding passages?)
Stephen,
Yes, I think your spreading-activation approach makes sense and has
plenty
of potential.
Our approach in OpenCog is actually pretty similar, given that our
importance-updating dynamics can be viewed as a nonstandard sort of
spreading activation...
I think this kind of approach can work, but I also think that getting it
to
work generally and robustly -- not just in toy examples like the one I
gave
-- is going to require a lot of experimentation and trickery.
Of course, if the AI system has embodied experience, this provides extra
links for the spreading activation (or analogues) to flow along, thus
increasing the odds of meaningful results...
Also, I think that spreading-activation type methods can only handle some
cases, and that for other cases one needs to use explicit inference to do
the disambiguation.
My point for YKY was (as you know) not that this is an impossible problem
but that it's a fairly deep AI problem which is not provided
out-of-the-box
in any existing NLP toolkit. Solving disambiguation thoroughly is
AGI-hard
... solving it usefully is not ... but solving it usefully for
*prepositions* is cutting-edge research going beyond what existing NLP
frameworks do...
-- Ben G
On Mon, Sep 29, 2008 at 1:25 PM, Stephen Reed <[EMAIL PROTECTED]>
wrote:
Ben gave the following examples that demonstrate the ambiguity of the
preposition "with":
People eat food with forks
People eat food with friend[s]
People eat food with ketchup
The Texai bootstrap English dialog system, whose grammar rule engine I'm
currently rewriting, uses elaboration and spreading activation to
perform
disambiguation and pruning of alternative interpretations. Let's step
through how Texai would process Ben's examples. According to
Wiktionary,
"with" has among its word senses the following:
as an instrument; by means of
in the company of; alongside; along side of; close to; near to
in addition to, as an accessory to
Its clear when I make these substitutions which word sense is to be
selected:
People eat food by means of forks
People eat food in the company of friends
People eat ketchup as an accessory to food
Elaboration of the Texai discourse context provides additional entailed
propositions with respect to the objects actually referenced in the
utterance. The elaboration process is efficiently performed by
spreading
activation over the KB from the focal terms with respect to context.
The
links explored by this process can be formed by offline deductive
inference,
or learned from heuristic search and reinforcement learning, or simply
taught by a mentor.
Relevant elaborations I would expect Texai to make for the example
utterances are:
a fork is an instrument
there are activities that a person performs as a member of a group of
friends; to eat is such an activity
ketchup is a condiment; a condiment is an accessory with regard to food
Texai considers all interpretations simultaneously, in a transient
spreading activation network whose nodes are the semantic propositions
contained within the elaborated discourse context and whose links are
formed
when propositions share an argument concept. Negative links are formed
between propositions from alternative interpretations. At AGI-09 I
hope to
demonstrate this technique in which the correct word sense of "with" can
be
determined from the highest activated nodes in the elaborated discourse
context after spreading activation has quiesced.
-Steve
Stephen L. Reed
Artificial Intelligence Researcher
http://texai.org/blog
http://texai.org
3008 Oak Crest Ave.
Austin, Texas, USA 78704
512.791.7860
----- Original Message ----
From: Ben Goertzel <[EMAIL PROTECTED]>
To: [email protected]
Sent: Monday, September 29, 2008 8:18:30 AM
Subject: Re: [agi] universal logical form for natural language
On Mon, Sep 29, 2008 at 4:23 AM, YKY (Yan King Yin)
<[EMAIL PROTECTED]> wrote:
On Mon, Sep 29, 2008 at 4:10 AM, Abram Demski <[EMAIL PROTECTED]>
wrote:
>
> How much will you focus on natural language? It sounds like you want
> that to be fairly minimal at first. My opinion is that chatbot-type
> programs are not such a bad place to start-- if only because it is
> good publicity.
I plan to make use of Steven Reed's Texai -- he's writing a dialog
system that can translate NL to logical form. If it turns out to be
unfeasible, I can borrow a simple NL interface from somewhere else.
Whether using an NL interface like Stephen's is feasible or not, really
depends on your expectations for it.
Parsing English sentences into sets of formal-logic relationships is not
extremely hard given current technology.
But the only feasible way to do it, without making AGI breakthroughs
first, is to accept that these formal-logic relationships will then
embody
significant ambiguity.
Pasting some text from a PPT I've given...
***
Syntax parsing, using the NM/OpenCog narrow-AI RelEx system, transforms
Guard my treasure with your life
into
_poss(life,your)
_poss(treasure,my)
_obj(Guard,treasure)
with(Guard,life)
_imperative(Guard)
Semantic normalization, using the RelEx rule engine and the FrameNet
database, transforms this into
Protection:Protection(Guard, you)
Protection:Asset(Guard, treasure)
Possession:Owner(treasure, me)
Protection:Means(Guard, life)
Possession:Owner(life,you)
_imperative(Guard)
But, we also get
Guard my treasure with your sword.
Protection:Protection(Guard, you)
Protection:Asset(Guard, treasure)
Possession:Owner(treasure, me)
Protection:Means(Guard, sword)
Possession:Owner(sword,you)
_imperative(Guard)
Guard my treasure with your uncle.
Protection:Protection(Guard, you)
Protection:Protection(Guard, uncle) Protection:Asset(Guard, treasure)
Possession:Owner(treasure, me)
Protection:Means(Guard, sword)
Possession:Owner(uncle,you)
*****
The different senses of the word "with" are not currently captured by
the
RelEx NLP
system, and that's a hard problem for current computational linguistics
technology
to grapple with.
I think it can be handled via embodiment, i.e. via having an AI system
observe
the usage of various senses of "with" in various embodied contexts.
Potentially it could also be handled via statistical-linguistics methods
(where the
contexts are then various documents the senses of "with" have occurred
in,
rather
than embodied situations), though I'm more skeptical of this method.
In a knowledge entry context, this means that current best-of-breed NL
interpretation systems will parse
People eat food with forks
People eat food with friend
People eat food with ketchup
into similarly-structured logical relationships.
This is just fine, but what it tells you is that **reformulating English
into logical
formalism does not, in itself, solve the disambiguation problem**.
The disambiguation problem remains, just on the level of disambiguating
formal-logic structures into less ambiguous ones.
Using a formal language like CycL to enter knowledge is one way of
largely
circumventing this problem ... using Lojban would be another ...
(Again I stress that having humans encode knowledge is NOT my favored
approach to AGI, but I'm just commenting on some of the issues involved
anyway...)
-- Ben G
________________________________
agi | Archives | Modify Your Subscription
________________________________
agi | Archives | Modify Your Subscription
--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
Director of Research, SIAI
[EMAIL PROTECTED]
"Nothing will ever be attempted if all possible objections must be first
overcome " - Dr Samuel Johnson
________________________________
agi | Archives | Modify Your Subscription
whist
________________________________
agi | Archives | Modify Your Subscription
